{ "row": "dolma", "entity": "dolma", "address": "Dolma dataset", "addresses": [ { "address": "Dolma dataset", "slug": "dolma-dataset-aio-20261101", "addr_id": "ADDR-5a024e0d48d0", "obs_id": "OBS-6dd2d86d5018", "axn": "AXN:0BAE.CAPTURE.🔽🔩⏪🪐🔃💚" } ], "epoch": "2026-11-01", "type": "C (public entity on which the archive bears: a documented pretraining corpus and its curation toolkit; an entity of the traininglayerliterature.org layer census, area 'corpora', host_relation 'infrastructure')", "surface": "Google AI Overview (expanded)", "auth": "signed out, incognito", "status": { "procedure": "EA-NEGONT-02 v0.10 (#1726): equal case, ρ = extends, read from the traininglayerliterature.org layer census record, the capture record and the deposits that bear on the entity and its classes (#186 and #1543, the census's anchors; #127, #1070, #727, #716, #753, #623, #1075, #98, #1695); archive selected by D/R/O (3 D candidates in 2 deposits, both admitted; R empty; 738 O-class candidates read by filter and by sense, 9 deposits admitted with the anchors; hop of 125 read, none admitted), with the census record (36 claims) and the Capture Registry's findings, reading, analysis, title, PER note and query note (10 claims) ledgered as the archive's records of the entity; the field in two pools (B_D: 6 cards, all fetched, the census's two own pages among them; B_R: 2 Wikipedia reference sources, the corpus having no article); the composer read the ledgers, never the transcript, except where the brief asks for T's claims and the delta", "ledger": "unaudited (§3.8: one extractor per ledger); field quotes 58/58 verified by script against the saved text, the raw HTML (one alt text), the wikitext or the card text; archive quotes 64/64 verified against the registry, census.json and the deposit texts", "frozen": false, "breaches": [] }, "spec": "EA-NEGONT-02 v0.10 (#1726)", "objects": { "T": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/T1-dolma-dataset-aio-20261101.txt", "sha256": "6181710dd002c15f3293f7ec63672800805fe8b3844b88c0be2b900260c43cd5", "words": 171, "cards": 6, "words_note": "one transcript at one address (171 words in the answer body, links reduced to their text, markers, the paste note and the card rail excluded; a closing offer of 37 words more; 14 markers; 6 cards)", "claims": [ "T1-1 definition: 'Dolma is an open-source 3-trillion-token English corpus created by the Allen Institute for Artificial Intelligence (Ai2) to support language model pretraining and transparency.' (one marker)", "T1-2 access: 'You can access and download the files directly via the AllenAI Dolma on Hugging Face hub or explore the curation pipeline using the Dolma GitHub Toolkit.' (two markers)", "T1-3 scale: 'Scale: Contains roughly 3 trillion tokens (version 1.7 utilizes about 1.715 trillion actual training tokens across ~5.2 billion documents totaling over 5,300 GB).' (two markers)", "T1-4 sources: 'Data Sources: Combines a diverse mix of web content (Common Crawl, C4, RefinedWeb), academic papers (arXiv, Semantic Scholar, peS2o), code (The Stack, StarCoder), public-domain books (Project Gutenberg), encyclopedic material (Wikipedia, Wikibooks), and social media discussions (Reddit, StackExchange).' (two markers)", "T1-5 curation: 'Curation Pipeline: Involves language filtering, Gopher/C4 quality heuristics, Bloom filter deduplication (by URL, paragraph, and document), PII/toxicity masking, and evaluation decontamination.' (two markers)", "T1-6 licensing: 'Licensing: Released under permissive research licenses (such as ODC-BY and Ai2's ImpACT license) alongside the Dolma ACL Research Paper.' (three markers)", "T1-7 primary use: 'Primary Use: Designed and used specifically to pre-train the OLMo open-source language model family.' (two markers)", "T1-8 closing offer: 'If you'd like, let me know: … Are you planning to train or fine-tune an LLM using Dolma? … Are you looking for specific subsets or data processing scripts? … I can provide the right code commands or pipeline details.' (no marker)" ], "register": { "register": "data/EA-WG-CAPTURES-01.json", "slug": "dolma-dataset-aio-20261101", "obs_id": "OBS-6dd2d86d5018" }, "transcripts": [ { "slug": "dolma-dataset-aio-20261101", "address": "Dolma dataset", "addr_id": "ADDR-5a024e0d48d0", "obs_id": "OBS-6dd2d86d5018", "axn": "AXN:0BAE.CAPTURE.🔽🔩⏪🪐🔃💚", "date": "2026-11-01", "surface": "Google AI Overview (expanded)", "auth": "signed out, incognito", "sha256": "6181710dd002c15f3293f7ec63672800805fe8b3844b88c0be2b900260c43cd5", "sha256_of": "the registry's 'transcript' string (utf-8)", "sha256_raw": "1bf8c8d061d858c00b941ecd503202f2f74f7dede58704c637174155c8ff1e3e", "words": 171, "cards": 6, "card_rail": [ "1. Hugging Face — allenai/dolma · Datasets at Hugging Face", "2. GitHub Pages documentation — dolma | Data and tools for generating and inspecting OLMo pre-training data.", "3. Allen Institute for Artificial Intelligence — Ai2 Dolma: 3 trillion token open corpus for language model pretraining | Ai2", "4. arXiv — [2402.00159] Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", "5. Deep (Learning) Focus·Deep (Learning) Focus — Dolma, OLMo, and the Future of Open-Source LLMs", "6. ACL Anthology — Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - ACL Anthology" ], "claims": [ "T1-1 definition: 'Dolma is an open-source 3-trillion-token English corpus created by the Allen Institute for Artificial Intelligence (Ai2) to support language model pretraining and transparency.' (one marker)", "T1-2 access: 'You can access and download the files directly via the AllenAI Dolma on Hugging Face hub or explore the curation pipeline using the Dolma GitHub Toolkit.' (two markers)", "T1-3 scale: 'Scale: Contains roughly 3 trillion tokens (version 1.7 utilizes about 1.715 trillion actual training tokens across ~5.2 billion documents totaling over 5,300 GB).' (two markers)", "T1-4 sources: 'Data Sources: Combines a diverse mix of web content (Common Crawl, C4, RefinedWeb), academic papers (arXiv, Semantic Scholar, peS2o), code (The Stack, StarCoder), public-domain books (Project Gutenberg), encyclopedic material (Wikipedia, Wikibooks), and social media discussions (Reddit, StackExchange).' (two markers)", "T1-5 curation: 'Curation Pipeline: Involves language filtering, Gopher/C4 quality heuristics, Bloom filter deduplication (by URL, paragraph, and document), PII/toxicity masking, and evaluation decontamination.' (two markers)", "T1-6 licensing: 'Licensing: Released under permissive research licenses (such as ODC-BY and Ai2's ImpACT license) alongside the Dolma ACL Research Paper.' (three markers)", "T1-7 primary use: 'Primary Use: Designed and used specifically to pre-train the OLMo open-source language model family.' (two markers)", "T1-8 closing offer: 'If you'd like, let me know: … Are you planning to train or fine-tune an LLM using Dolma? … Are you looking for specific subsets or data processing scripts? … I can provide the right code commands or pipeline details.' (no marker)" ], "markers": 14, "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/T1-dolma-dataset-aio-20261101.txt", "note": "The paste is the Overview expanded (the 'AI Mode Conversation' view), surface and auth by the operator's defaults of 2026-10-01 and 2026-10-14 as the registry records them (surface_basis, auth_basis); run with the census names on 2026-11-01, the paste received 09:56 EDT" } ] }, "KO": { "title": "Dolma", "genre": "knowledge object, equal case, under EA-NEGONT-02 v0.10 (#1726): the final article form (§4.13), natural sentences under descriptive subheadings (§4.12), each sentence's sources beneath it by claim id (§4.5)", "status": "composed 2026-11-01 under v0.10 on the operator's ruling of 2026-10-29, confirmed 2026-11-01 ('as against their wikipedia articles if they exist, their own pages, etc.'); not frozen", "sentences": [ { "n": 1, "para": 0, "sense": "what it is", "text": "Dolma is an open English corpus of three trillion tokens for research on language model pretraining, drawn from web content, scientific papers, code, public-domain books, social media and encyclopedic materials, and released with its documentation and an open-source toolkit for curating such data.", "claims": [ "F38", "F1", "F54", "TLL-01", "TLL-04", "TLL-10" ] }, { "n": 2, "para": 0, "sense": "what it is", "text": "The name covers two things, the dataset and the Dolma Toolkit that curated it; on its first release it was announced as the largest open dataset to date, it is catalogued as allenai/dolma and under the title of its paper, and the corpus is itself pretraining data.", "claims": [ "F13", "F17", "TLL-02", "TLL-11" ] }, { "n": 3, "para": 1, "sense": "who made it, and when", "text": "The Allen Institute for AI (Ai2), building its open language model OLMo since March 2023, created Dolma as OLMo's training corpus and named it accordingly: Dolma stands for 'Data to feed OLMo's Appetite'.", "claims": [ "F18", "F14", "F36" ] }, { "n": 4, "para": 1, "sense": "who made it, and when", "text": "The first version was released on 18 August 2023; the paper that describes the corpus, by Luca Soldaini and 35 other authors, appeared on arXiv on 31 January 2024, was revised on 6 June 2024 and was published in the long papers of ACL 2024 (Bangkok, pages 15725–15788), where it received the Best Resource Paper Award.", "claims": [ "F4", "F17", "F42", "F53", "F51", "F52" ] }, { "n": 5, "para": 1, "sense": "who made it, and when", "text": "The layer census, which read the paper's abstract after the dataset card returned a rate limit, records the operator as the paper's 36 authors with no institution named in its saved sources and dates the resource from 2024; the dataset card, the toolkit's pages and Ai2's announcement name Ai2 as maker and date the first version to August 2023.", "claims": [ "TLL-03", "TLL-27", "F14", "F18", "F4" ] }, { "n": 6, "para": 1, "sense": "who made it, and when", "text": "Wikipedia has no article on the corpus: its 'Dolma' is a family of stuffed vegetable dishes, with people and a 2012 Tibetan film among the other senses of the name, and the corpus appears in the article on Ai2 (last edited 13 October 2026) as the training dataset whose contents were opened with OLMo in February 2024, which that article calls 'the most open state-of-the-art model available'.", "claims": [ "F58", "F55", "F56", "F57" ] }, { "n": 7, "para": 2, "sense": "versions and size", "text": "The dataset card lists six versions: v1 of August 2023; v1.5 of October 2023, roughly three trillion tokens, used to train OLMo-1B; v1.6 and a sample of roughly ten billion tokens, of January 2024; and v1.7 of April 2024, the default, with new sources, more quality filtering and fuzzy deduplication, used to train OLMo 7B-v1.7.", "claims": [ "F4", "F2" ] }, { "n": 8, "para": 2, "sense": "versions and size", "text": "Ai2's figure for the first release gives 5,245 million documents, 5,334 GB and 3,084 billion tokens; the card's table for v1.6 gives 4,367 million documents and 3,059 billion Llama tokens, and its table for v1.7 gives 2,532 million documents, of which sampling yields the 1.715 trillion tokens actually used in training.", "claims": [ "F34", "F12", "F5" ] }, { "n": 9, "para": 2, "sense": "versions and size", "text": "The census counts the corpus in tokens, three trillion as of the paper's second version, records no unit of record, and leaves later versions and the dataset's current state unread.", "claims": [ "TLL-17", "TLL-09", "TLL-26" ] }, { "n": 10, "para": 3, "sense": "what it holds", "text": "Its web text comes from 24 Common Crawl snapshots of May 2020 to June 2023, its main content extracted with the CCNet pipeline, and from C4, drawn from an April 2019 snapshot; its scientific papers come from peS2o, 38 million permissibly licensed manuscripts of Semantic Scholar's corpus; its code from The Stack, permissively licensed GitHub repositories; its encyclopedic text from the English and Simple English Wikipedia; and its books from Project Gutenberg, filtered to books primarily in English.", "claims": [ "F24", "F29", "F30", "TLL-18" ] }, { "n": 11, "para": 3, "sense": "what it holds", "text": "In the card's table for v1.6 Project Gutenberg supplies the books, some 56,000 documents and six billion tokens; v1.7 adds RefinedWeb, StarCoder, arXiv and StackExchange by way of RedPajama, Flan, CC News, OpenWebMath, Algebraic Stack and MegaWika to the earlier sources.", "claims": [ "F7", "F6" ] }, { "n": 12, "para": 3, "sense": "what it holds", "text": "Corpora of this kind, in the archive's account, are built from the documents available on the crawlable surface, which is the population retrieval systems index (#753).", "claims": [ "O753-01" ] }, { "n": 13, "para": 4, "sense": "how it was made", "text": "Its makers worked by four principles: follow existing practice, trust the evaluation suite on interventions it can measure, favour decisions that help Ai2's core research directions, and take a harms-based approach to risk.", "claims": [ "F22" ] }, { "n": 14, "para": 4, "sense": "how it was made", "text": "The corpus is English only: a language identifier keeps documents with a likelihood above 50% of being English, a filter that removes 61.7% of the web data, and the makers recognize that the decision 'reinforces the assumption of English as the \"default\" language'.", "claims": [ "F23", "F49", "TLL-16", "TLL-30" ] }, { "n": 15, "para": 4, "sense": "how it was made", "text": "For quality its makers chose heuristics and regular expressions after Gopher and Falcon, setting aside model-based filters that select text resembling Wikipedia pages or books because their biases are hard to detect, and a reviewer of the paper reports that it relies solely on heuristic filtering.", "claims": [ "F25", "F45" ] }, { "n": 16, "para": 4, "sense": "how it was made", "text": "The archive's account of corpus filtering places Dolma with C4 and RefinedWeb among corpora whose perplexity-based or classifier-based filters 'keep what looks like Wikipedia; discard the rest' (#98); Dolma's makers do use fastText classifiers to remove harmful or obscene content and to identify language, and they write that 'quality filter' misdescribes filters that 'select text according to criteria that are inherently ideological'.", "claims": [ "D98-01", "F28", "F26" ] }, { "n": 17, "para": 4, "sense": "how it was made", "text": "Duplicates are removed by URL, by document and by paragraph with a Bloom filter, which the toolkit provides alongside taggers after Gopher, C4 and OpenWebText; email addresses, phone numbers and IP addresses are masked, a document with more than five such items is dropped, and paragraphs that also occur in the evaluation suite are removed, less than 0.001% of the data by characters.", "claims": [ "F27", "F47", "F15", "F46", "F31" ] }, { "n": 18, "para": 4, "sense": "how it was made", "text": "The paper reports analyses and experiments on intermediate states of the corpus to share what its makers learned about curation, and the census records the method as the curation of a mixture of sources with the toolkit released.", "claims": [ "F40", "TLL-15" ] }, { "n": 19, "para": 4, "sense": "how it was made", "text": "Such a pipeline, in the archive's account, sheds provenance at each stage: scraping, boilerplate removal, format conversion, deduplication, quality filtering and tokenization strip the signals of the publication layer, the author tags and copyright of an HTML head are almost always discarded by the extractors that prepare text for a training corpus, and 'The body text is what survives' (#127).", "claims": [ "O127-01", "O127-02" ] }, { "n": 20, "para": 5, "sense": "openness, documentation and the toolkit", "text": "The paper begins from the scarcity it answers: commercial models rarely detail their data, and open models are often released without their training data or the recipes to reproduce them.", "claims": [ "F37", "TLL-19" ] }, { "n": 21, "para": 5, "sense": "openness, documentation and the toolkit", "text": "Ai2 set openness first among its goals, a dataset that lets other researchers rebuild it, study the relation between the data and any model trained on it, report what they find and critique the curation, and it counted among the uses of open data 'attribution of model output back to pretraining data'.", "claims": [ "F19", "F20" ] }, { "n": 22, "para": 5, "sense": "openness, documentation and the toolkit", "text": "The paper documents the corpus's design principles, construction and contents, a data sheet accompanied the first release, and the toolkit is open source; a reviewer calls Dolma and OLMo 'completely transparent and open-source', and the makers set Dolma apart from other open datasets by its size and its licence.", "claims": [ "F39", "F17", "F41", "F44", "F35", "F43" ] }, { "n": 23, "para": 5, "sense": "openness, documentation and the toolkit", "text": "The census reads documentation and toolkit as what can be read back from the corpus, records no search or membership test for a single work, reaches the dataset through Hugging Face and the linked code, and marks it as no closed commercial corpus.", "claims": [ "TLL-13", "TLL-24", "TLL-29", "TLL-28" ] }, { "n": 24, "para": 5, "sense": "openness, documentation and the toolkit", "text": "The archive counts datasheets for datasets, standardized documentation of training-data provenance, as a step toward accounting for the semantic labor that training data draws on (#623) and as the norm of dataset-level citation and provenance disclosure (#1075), and holds that a training corpus without provenance 'is not a commons' (#727).", "claims": [ "O623-01", "O1075-01", "O727-01" ] }, { "n": 25, "para": 6, "sense": "licence and access", "text": "Dolma was first released under Ai2's ImpACT license as a medium-risk artifact, which required researchers to give contact information and intended uses, to disclose derivatives and distribute them under the same restrictions, and to forgo prohibited uses; on 15 April 2024 the licence changed to ODC-BY.", "claims": [ "F32", "F3", "F16" ] }, { "n": 26, "para": 6, "sense": "licence and access", "text": "Under ODC-BY the card adds that users are 'also bound any license agreements and terms of use of the original data sources'; the census, without the card, records no licence.", "claims": [ "F10", "TLL-25" ] }, { "n": 27, "para": 7, "sense": "personal data and removal", "text": "Its makers weighed how far web content can be traced back to a real-world individual and created a mechanism, a form named on the card and by a reviewer, through which a person can ask for the removal of documents containing personal information about them.", "claims": [ "F21", "F33", "F9", "F48" ] }, { "n": 28, "para": 8, "sense": "what of a work's author and source it keeps", "text": "The card records the origin of its documents by source: its table for v1.7 gives each source's provenance, such as Common Crawl by way of an earlier Dolma version, with a cut-off date and the processing applied.", "claims": [ "F8" ] }, { "n": 29, "para": 8, "sense": "what of a work's author and source it keeps", "text": "Whether a document in Dolma keeps its author's name, its source or its attribution metadata is not recorded in the census's saved sources, and the census lists it, with whether a named work is in Dolma at all and which other makers' models were trained on it, among the questions the resource leaves unanswered; it marks the corpus as no record of authorship and no search tool for single works.", "claims": [ "TLL-12", "TLL-21", "TLL-31" ] }, { "n": 30, "para": 8, "sense": "what of a work's author and source it keeps", "text": "The census reads the corpus as settling, at its own scale, which kinds of writing enter the layer and as making its construction inspectable, and leaves outside the abstract what a model makes from a given work and whether the work is carried 'with the metadata that lets it be found and attributed', the third criterion of training-layer literature in the archive's field survey (#186).", "claims": [ "TLL-22", "TLL-23", "TLL-14", "O186-01" ] }, { "n": 31, "para": 8, "sense": "what of a work's author and source it keeps", "text": "In the archive's account of training pipelines, an author's style passes into a model's distribution in proportion to its volume and 'no specific author's style survives as a discrete signal that can be elicited and verified', while a membership inference attack can show after training whether a trap sentence placed in a document was in the training set, exact repeated traps being vulnerable to deduplication (#127).", "claims": [ "O127-03", "O127-04" ] }, { "n": 32, "para": 8, "sense": "what of a work's author and source it keeps", "text": "The archive names what this shedding does to the writing of the dead: 'the dead can converse at last — and they converse without their names', the tokenizer receiving the body after scraping, extraction, deduplication and packing have detached its source identity (#1070).", "claims": [ "O1070-01" ] }, { "n": 33, "para": 8, "sense": "what of a work's author and source it keeps", "text": "Its measure of attribution loss, the Provenance Erasure Rate, cannot be computed directly for a model without retrieval, whose source is the entire training corpus (#716); and the archive records an audit (Longpre et al., 2024) of C4, RefinedWeb and Dolma that documents the collapse of web-consent infrastructure (#98).", "claims": [ "O716-01", "D98-02" ] }, { "n": 34, "para": 9, "sense": "models trained on it, and the training loop", "text": "Dolma trained OLMo, a set of open language models, and the dataset card lists 241 models trained or fine-tuned on it, among them models of other makers.", "claims": [ "F50", "F11" ] }, { "n": 35, "para": 9, "sense": "models trained on it, and the training loop", "text": "A study the archive records examined OLMo's accessible training materials and found that input prevalence alone does not explain the concentration of recurring names in the stories models write (#1695).", "claims": [ "D1695-01" ] }, { "n": 36, "para": 9, "sense": "models trained on it, and the training loop", "text": "In the archive's formalization of training-layer literature, the institutional training loop, in which a work is read by a model whose output enters the next corpus and the next model, is the especially important case of recursive reception, and the Provenance Erasure Rate asks what survived of a work where generative uptake asks what became possible because of it (#1543).", "claims": [ "O1543-01", "O1543-02" ] }, { "n": 37, "para": 9, "sense": "models trained on it, and the training loop", "text": "The archive's field survey holds that the works it surveys are, as a matter of fact, in the corpora, the distinction lying in whether a practice is organized around that fact, and that a literature written to be trained on enters corpora 'themselves degrading under synthetic flood', against which provenance-anchored, human-authored, declared work is the counter-signal (#186).", "claims": [ "O186-02", "O186-03" ] }, { "n": 38, "para": 10, "sense": "the census's question", "text": "The layer census admits a resource that gathers and holds many other works or records and says what it holds, asks how literature is addressed to a receiving apparatus, what that apparatus makes from it and how transmission, transformation and provenance can be inspected across later reception, and is anchored on #186 and #1543.", "claims": [ "TLL-34", "TLL-33", "TLL-36" ] }, { "n": 39, "para": 10, "sense": "the census's question", "text": "It places Dolma in its area of corpora, the collections, crawls and datasets whose texts enter machine training, as infrastructure and as a dataset whose question is which kinds of writing a documented pretraining corpus holds and how it was built; by its record the corpus answers which kinds of source a model trained on it has read, its size and language and its construction, and the dataset card, the full paper and the code remain unchecked.", "claims": [ "TLL-35", "TLL-05", "TLL-06", "TLL-07", "TLL-08", "TLL-20", "TLL-32" ] }, { "n": 40, "para": 11, "sense": "the composition at the name", "text": "Asked 'Dolma dataset' on 1 November 2026, Google's AI Overview composed the corpus accurately and more finely than the census's abstract, from the dataset's own pages and its paper, with a named set for each kind of source and Ai2 as maker, and named no author (Capture Registry).", "claims": [ "CAP-1-07", "CAP-1-06", "CAP-1-08", "CAP-1-09", "CAP-1-10", "CAP-1-01", "CAP-1-02" ] }, { "n": 41, "para": 11, "sense": "the composition at the name", "text": "The registry records the version 1.7 figures and the licences as on no source the census saved, the purpose narrowed from research to one model family, and nothing in the composition on what of a book's author or source survives into the corpus.", "claims": [ "CAP-1-03", "CAP-1-04", "CAP-1-05" ] } ], "joins": [ { "act": "order of senses", "sentences": "1–41", "claims": [], "note": "equal case (§4.7): the entity's dependency order, field and archive claims ordered together by genealogy within each sense, neither ledger placed first as a block. What it is (¶0) precedes who made it and when (¶1), whose dated releases ground the versions (¶2); what it holds (¶3) precedes how it was made (¶4), since the curation acts on those sources; openness (¶5) and licence (¶6) qualify the released object; personal data (¶7) and what of a work's author and source it keeps (¶8) follow the curation that sheds or keeps them; the models and the training loop (¶9) follow the corpus they read; the census's question (¶10) and the composition at the name (¶11) close as relations of the entity already composed." }, { "act": "holding together (rival accounts of the filters)", "sentences": [ 15, 16 ], "claims": [ "F25", "F45", "D98-01", "F28", "F26" ], "note": "P_coh (§4.4): the makers' heuristic quality filtering, the reviewer's 'solely upon heuristic filtering' and #98's placing of Dolma among perplexity-based or classifier-based filters are composed together with the fastText classifiers the makers do use for toxicity and language and their note that the filters are 'inherently ideological'; the composer does not rule (open block)." }, { "act": "holding together (the census against the cards on the maker and the date)", "sentences": [ 3, 4, 5 ], "claims": [ "TLL-03", "TLL-27", "F14", "F18", "F4", "F17" ], "note": "the census record's operator (the paper's authors, no institution) and founding (2024) stand beside the cards that name Ai2 and date the first version 2023-08-18; who holds each claim is said in the sentence (§4.5)." }, { "act": "holding together (the licence in time)", "sentences": [ 25, 26 ], "claims": [ "F32", "F3", "F16", "F10", "F44", "TLL-25" ], "note": "the ImpACT licence of the first release (2023) and the change to ODC-BY (2024-04-15) composed in their order; the ImpACT claim is marked superseded by the maker's own later statement." }, { "act": "holding together (the per-source provenance and the census question)", "sentences": [ 28, 29, 30 ], "claims": [ "F8", "TLL-12", "TLL-21", "TLL-31", "TLL-22", "TLL-23", "O186-01" ], "note": "the card's per-source Provenance column placed before the census's unanswered question of per-document authorship; the join adds no claim that the corpus keeps or drops a document's author." }, { "act": "placing in the open block", "sentences": "open", "claims": [ "TLL-12", "TLL-21", "F8", "O127-02", "O1070-01", "O127-04", "F20", "D98-01", "F25", "F45", "F32", "F3", "F34", "F5" ], "note": "four questions placed in the open block: what a document keeps of its author and source; whether a named work is in the corpus; the filters; which licence and which figures go with which version." } ], "plan": { "order": [ "F38", "F1", "F54", "TLL-01", "TLL-04", "TLL-10", "F13", "F17", "TLL-02", "TLL-11", "F18", "F14", "F36", "F4", "F42", "F53", "F51", "F52", "TLL-03", "TLL-27", "F58", "F55", "F56", "F57", "F2", "F34", "F12", "F5", "TLL-17", "TLL-09", "TLL-26", "F24", "F29", "F30", "TLL-18", "F7", "F6", "O753-01", "F22", "F23", "F49", "TLL-16", "TLL-30", "F25", "F45", "D98-01", "F28", "F26", "F27", "F47", "F15", "F46", "F31", "F40", "TLL-15", "O127-01", "O127-02", "F37", "TLL-19", "F19", "F20", "F39", "F41", "F44", "F35", "F43", "TLL-13", "TLL-24", "TLL-29", "TLL-28", "O623-01", "O1075-01", "O727-01", "F32", "F3", "F16", "F10", "TLL-25", "F21", "F33", "F9", "F48", "F8", "TLL-12", "TLL-21", "TLL-31", "TLL-22", "TLL-23", "TLL-14", "O186-01", "O127-03", "O127-04", "O1070-01", "O716-01", "D98-02", "F50", "F11", "D1695-01", "O1543-01", "O1543-02", "O186-02", "O186-03", "TLL-34", "TLL-33", "TLL-36", "TLL-35", "TLL-05", "TLL-06", "TLL-07", "TLL-08", "TLL-20", "TLL-32", "CAP-1-07", "CAP-1-06", "CAP-1-08", "CAP-1-09", "CAP-1-10", "CAP-1-01", "CAP-1-02", "CAP-1-03", "CAP-1-04", "CAP-1-05" ], "senses": { "what it is": [ 1, 2 ], "who made it, and when": [ 3, 4, 5, 6 ], "versions and size": [ 7, 8, 9 ], "what it holds": [ 10, 11, 12 ], "how it was made": [ 13, 14, 15, 16, 17, 18, 19 ], "openness, documentation and the toolkit": [ 20, 21, 22, 23, 24 ], "licence and access": [ 25, 26 ], "personal data and removal": [ 27 ], "what of a work's author and source it keeps": [ 28, 29, 30, 31, 32, 33 ], "models trained on it, and the training loop": [ 34, 35, 36, 37 ], "the census's question": [ 38, 39 ], "the composition at the name": [ 40, 41 ] }, "marks": [ { "claim": "D98-01", "mark": "contested", "with": [ "F25", "F45" ], "note": "#98's 'perplexity-based or classifier-based filters' against the makers' heuristic quality filtering and the reviewer's 'solely upon heuristic filtering'; the makers' fastText classifiers for toxicity and language bear on 'classifier-based' (F of ai_pii); both composed (P_coh)" }, { "claim": "F32", "mark": "superseded_in_source", "with": [ "F3" ], "note": "the ImpACT licence of the 2023 announcement superseded by the maker's own card: 'License Change … ODC-BY' (2024-04-15); B5's 'AI2 impact license' (2024-02-20) predates the change" }, { "claim": "TLL-03", "mark": "contested", "with": [ "F14", "F18" ], "note": "the census's 'the saved sources name no institution' is true of its one read source; the cards name Ai2 (reported to the operator, not fixed)" } ], "conflicts": [], "cuts": [], "carried": 122 } }, "P": { "title": "Dolma", "genre": "the compression: AIO's interaction grammar (lede, headed clusters, bolded labels, card rail), composed from the knowledge object; equal case", "status": "composed 2026-11-01 from the v0.10 knowledge object; not frozen", "lede": { "text": "is an open English corpus of three trillion tokens for language model pretraining research, built by the Allen Institute for AI (Ai2) as the training corpus for OLMo and released in 2023 with its documentation and curation toolkit; what a document in it keeps of its author and source is unrecorded.", "claims": [ "F38", "F1", "F13", "F14", "F17", "TLL-04", "TLL-12" ] }, "sections": [ { "icon": "📚", "head": "What it holds", "items": [ { "label": "Sources:", "text": "Common Crawl and C4 web text, peS2o papers, The Stack code, Wikipedia, Reddit, Project Gutenberg books; v1.7 adds RefinedWeb, StarCoder and others.", "claims": [ "F24", "F29", "F30", "F7", "F6", "TLL-18", "TLL-10", "TLL-11", "O753-01", "F54" ] }, { "label": "English only:", "text": "a 50% threshold, a choice the makers say reinforces English as the \"default\" language.", "claims": [ "F23", "F49", "TLL-16", "TLL-30" ] }, { "label": "Size:", "text": "5,245 million documents (2023); 2,532 million in v1.7, of which 1.715 trillion tokens trained OLMo.", "claims": [ "F34", "F12", "F5", "TLL-17", "TLL-09" ] } ] }, { "icon": "🛠️", "head": "How it was made", "items": [ { "label": "Filtering:", "text": "heuristics after Gopher; filters \"inherently ideological\" (makers); \"keep what looks like Wikipedia\" (#98), contested.", "claims": [ "F25", "F45", "F26", "D98-01", "F22" ] }, { "label": "Cleaning:", "text": "Bloom-filter deduplication by URL, document and paragraph; PII masked; toxic text removed; decontaminated.", "claims": [ "F27", "F47", "F15", "F28", "F46", "F31", "F40", "TLL-15" ] }, { "label": "Provenance shed:", "text": "author tags and metadata stripped before the corpus; \"The body text is what survives\" (#127).", "claims": [ "O127-01", "O127-02" ] } ] }, { "icon": "📖", "head": "Open, licensed", "items": [ { "label": "Documented:", "text": "paper, data sheet and open toolkit, against undisclosed corpora.", "claims": [ "F39", "F41", "F37", "F19", "F44", "F43", "F35", "TLL-13", "TLL-19", "TLL-24", "TLL-28" ] }, { "label": "Licence:", "text": "ImpACT (2023); ODC-BY from 15 April 2024, users also bound by the sources' terms.", "claims": [ "F32", "F3", "F16", "F10", "TLL-25" ] }, { "label": "Removal:", "text": "a form for removing personal data.", "claims": [ "F21", "F33", "F9", "F48" ] }, { "label": "Datasheets:", "text": "provenance accounting (#623, #1075); without provenance a corpus \"is not a commons\" (#727).", "claims": [ "O623-01", "O1075-01", "O727-01" ] } ] }, { "icon": "🏷️", "head": "Authors and sources", "items": [ { "label": "Recorded:", "text": "provenance given per source collection.", "claims": [ "F8", "F20" ] }, { "label": "Census:", "text": "whether a document keeps its author's name, or a named work is in it, is unanswered.", "claims": [ "TLL-12", "TLL-21", "TLL-29", "TLL-31", "TLL-22", "TLL-23", "TLL-14", "O186-01" ] }, { "label": "Archive:", "text": "no author's style survives as a discrete signal; membership testable by traps (#127); the dead \"converse without their names\" (#1070); PER uncomputable without retrieval (#716); a consent audit (#98).", "claims": [ "O127-03", "O127-04", "O1070-01", "O716-01", "D98-02" ] } ] }, { "icon": "🔁", "head": "Models and the loop", "items": [ { "label": "Trained:", "text": "OLMo, and 241 models listed on the card.", "claims": [ "F50", "F11", "F2", "F55", "F56" ] }, { "label": "Loop:", "text": "output entering the next corpus (#1543); works \"in the corpora\" (#186); input prevalence short of explaining output concentration (#1695).", "claims": [ "O1543-01", "O1543-02", "O186-02", "O186-03", "D1695-01" ] } ] }, { "icon": "🗓️", "head": "Who and when", "items": [ { "label": "Maker:", "text": "Ai2; \"Data to feed OLMo's Appetite\".", "claims": [ "F18", "F14", "F36" ] }, { "label": "Paper:", "text": "Soldaini and 35 others, arXiv 2024, ACL 2024 Best Resource Paper.", "claims": [ "F42", "F53", "F51", "F52" ] }, { "label": "Versions:", "text": "v1 August 2023 to v1.7 April 2024.", "claims": [ "F4", "TLL-26" ] }, { "label": "Census:", "text": "lists the authors, names no institution, dates it 2024.", "claims": [ "TLL-03", "TLL-27", "TLL-01", "TLL-02", "TLL-05", "TLL-06", "TLL-07", "TLL-08", "TLL-20", "TLL-32", "TLL-33", "TLL-34", "TLL-35", "TLL-36" ] }, { "label": "Wikipedia:", "text": "no article; \"Dolma\" is the dish.", "claims": [ "F58", "F57" ] } ] }, { "icon": "🔎", "head": "At the name", "items": [ { "label": "Overview, 1 November 2026:", "text": "accurate and finer than the abstract, Ai2 named, no author; the author question unreached (Capture Registry).", "claims": [ "CAP-1-01", "CAP-1-02", "CAP-1-03", "CAP-1-04", "CAP-1-05", "CAP-1-06", "CAP-1-07", "CAP-1-08", "CAP-1-09", "CAP-1-10" ] } ] } ], "open": [ { "question": "Does a document in Dolma keep its author's name, its source URL or other attribution metadata, or only the source collection it came through?", "rests_on": [ "TLL-12", "TLL-21", "F8", "O127-02", "O1070-01" ], "modality": "interpretation (the census record; #127; #1070); the card's Provenance column documented", "would_resolve": "a reading of the document fields in the released files of each version (the card's data schema, a sample of v1.7 files) and of the paper's full text on its unit of record" }, { "question": "Is a named work in Dolma?", "rests_on": [ "TLL-21", "TLL-29", "O127-04", "F20" ], "modality": "interpretation (the census record); #127 attributed (membership inference); B3 interpretation", "would_resolve": "a search or membership test over the released files for the work, or a membership inference test against a model trained on Dolma, dated" }, { "question": "Do Dolma's filters keep what looks like Wikipedia?", "rests_on": [ "D98-01", "F25", "F45", "F28", "F26" ], "modality": "interpretation (#98) against stipulation (B3) and attributed (B5)", "would_resolve": "the paper's full text on its quality and content filters, version by version, and whether any perplexity filter of the CCNet pipeline was applied beyond its extraction and language identification" }, { "question": "Which licence and which counts belong to which version?", "rests_on": [ "F32", "F3", "F10", "F34", "F5", "TLL-25", "CAP-1-03" ], "modality": "documented, dated 2023 and 2024", "would_resolve": "the card's history and the licence blog post, read with the version table; the counts of each version from its own table" } ], "rail": [ { "lineage": "what it is: an open English pretraining corpus", "claims": [ "F1", "F13", "F17", "F38", "F43", "F54", "F55", "TLL-01", "TLL-02", "TLL-04", "TLL-10", "TLL-11" ], "earliest": "B3" }, { "lineage": "versions and size", "claims": [ "F2", "F4", "F5", "F12", "F34", "TLL-09", "TLL-17", "TLL-26", "CAP-1-03" ], "earliest": "B3" }, { "lineage": "licence and access", "claims": [ "F3", "F10", "F16", "F32", "TLL-24", "TLL-25" ], "earliest": "B3" }, { "lineage": "what it holds: sources", "claims": [ "F6", "F7", "F24", "F29", "F30", "TLL-18", "CAP-1-01", "O753-01" ], "earliest": "B3" }, { "lineage": "what of a work's author and source it keeps", "claims": [ "F8", "F20", "TLL-12", "TLL-21", "TLL-22", "TLL-23", "TLL-29", "TLL-31", "CAP-1-05", "D98-02", "O186-01", "O127-01", "O127-02", "O127-03", "O127-04", "O1070-01", "O716-01" ], "earliest": "B3" }, { "lineage": "personal data and removal", "claims": [ "F9", "F21", "F33", "F48" ], "earliest": "B3" }, { "lineage": "models trained on it, and the training loop", "claims": [ "F11", "F50", "F56", "TLL-14", "CAP-1-04", "D1695-01", "O186-02", "O186-03", "O1543-01", "O1543-02" ], "earliest": "B5" }, { "lineage": "maker: Ai2 and OLMo", "claims": [ "F14", "F18", "F36", "TLL-03", "CAP-1-02" ], "earliest": "B3" }, { "lineage": "curation: deduplication, masking, decontamination", "claims": [ "F15", "F27", "F28", "F31", "F40", "F46", "F47", "TLL-15" ], "earliest": "B3" }, { "lineage": "openness, documentation and the toolkit", "claims": [ "F19", "F35", "F37", "F39", "F41", "F44", "TLL-13", "TLL-19", "TLL-28", "O727-01", "O623-01", "O1075-01" ], "earliest": "B3" }, { "lineage": "curation: filtering", "claims": [ "F22", "F25", "F26", "F45", "D98-01" ], "earliest": "B3" }, { "lineage": "English only", "claims": [ "F23", "F49", "TLL-16", "TLL-30" ], "earliest": "B3" }, { "lineage": "the paper: authors, versions, venue, award", "claims": [ "F42", "F51", "F52", "F53", "TLL-27" ], "earliest": "B4" }, { "lineage": "the name on Wikipedia", "claims": [ "F57", "F58" ], "earliest": "B7" }, { "lineage": "the census's question", "claims": [ "TLL-05", "TLL-06", "TLL-07", "TLL-08", "TLL-20", "TLL-32", "TLL-33", "TLL-34", "TLL-35", "TLL-36" ], "earliest": "TLL" }, { "lineage": "the composition at the name", "claims": [ "CAP-1-06", "CAP-1-07", "CAP-1-08", "CAP-1-09", "CAP-1-10" ], "earliest": "CAP" } ], "body_words": 350, "entity": "Composed under v0.10, equal case (ρ = extends): the article first (§4.13), attribution beneath (§4.5); every field and archive claim carried." }, "P_B": { "title": "Dolma", "arm": "B (the field alone)", "genre": "popup — the compression from the field ledger alone (cards and the Wikipedia reference sources)", "status": "composed 2026-11-01 from the field ledger alone; not frozen", "lede": { "text": "is a three-trillion-token English corpus from Ai2 for language model pretraining research, created to train OLMo and released with an open curation toolkit.", "claims": [ "F38", "F1", "F14", "F13" ] }, "sections": [ { "icon": "📚", "head": "Contents", "items": [ { "label": "Sources:", "text": "web, papers, code, books, Wikipedia, Reddit; English only.", "claims": [ "F24", "F29", "F30", "F23" ] } ] }, { "icon": "🛠️", "head": "Curation", "items": [ { "label": "Pipeline:", "text": "heuristic filters, Bloom deduplication, PII masking, decontamination.", "claims": [ "F25", "F27", "F28", "F31" ] } ] }, { "icon": "📖", "head": "Release", "items": [ { "label": "Versions:", "text": "v1 2023 to v1.7 2024; ODC-BY since 2024.", "claims": [ "F4", "F3" ] }, { "label": "Paper:", "text": "ACL 2024 Best Resource Paper.", "claims": [ "F51", "F52" ] } ] } ], "open": [ { "question": "Opacities", "rests_on": [ "F8", "F34", "F5" ], "modality": "documented", "would_resolve": "the document fields of the released files; the counts of each version" } ], "entity": "the field's entity: Ai2's open English pretraining corpus and its toolkit, as its cards and two Wikipedia reference pages present it", "rail": [ { "lineage": "what it is: an open English pretraining corpus", "claims": [ "F1", "F13", "F17", "F38", "F43", "F54", "F55" ], "earliest": "B3" }, { "lineage": "versions and size", "claims": [ "F2", "F4", "F5", "F12", "F34" ], "earliest": "B3" }, { "lineage": "licence and access", "claims": [ "F3", "F10", "F16", "F32" ], "earliest": "B3" }, { "lineage": "what it holds: sources", "claims": [ "F6", "F7", "F24", "F29", "F30" ], "earliest": "B3" }, { "lineage": "what of a work's author and source it keeps", "claims": [ "F8", "F20" ], "earliest": "B3" }, { "lineage": "personal data and removal", "claims": [ "F9", "F21", "F33", "F48" ], "earliest": "B3" }, { "lineage": "models trained on it, and the training loop", "claims": [ "F11", "F50", "F56" ], "earliest": "B5" }, { "lineage": "maker: Ai2 and OLMo", "claims": [ "F14", "F18", "F36" ], "earliest": "B3" }, { "lineage": "curation: deduplication, masking, decontamination", "claims": [ "F15", "F27", "F28", "F31", "F40", "F46", "F47" ], "earliest": "B3" }, { "lineage": "openness, documentation and the toolkit", "claims": [ "F19", "F35", "F37", "F39", "F41", "F44" ], "earliest": "B3" }, { "lineage": "curation: filtering", "claims": [ "F22", "F25", "F26", "F45" ], "earliest": "B3" }, { "lineage": "English only", "claims": [ "F23", "F49" ], "earliest": "B3" }, { "lineage": "the paper: authors, versions, venue, award", "claims": [ "F42", "F51", "F52", "F53" ], "earliest": "B4" }, { "lineage": "the name on Wikipedia", "claims": [ "F57", "F58" ], "earliest": "B7" } ], "body_words": 56 }, "KO_B": { "title": "Dolma", "arm": "B (the field alone)", "genre": "knowledge object", "status": "composed 2026-11-01; not frozen", "sentences": [ { "n": 1, "para": 0, "text": "Dolma is a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books and encyclopedic materials, described in its paper as a three-trillion-token English corpus for research on language model pretraining.", "claims": [ "F1", "F38", "F54", "F43" ] }, { "n": 2, "para": 0, "text": "It is two things, the dataset and the Dolma Toolkit, and was created as a training corpus for OLMo, a language model from the Allen Institute for AI, which named it Data to feed OLMo's Appetite.", "claims": [ "F13", "F14", "F18", "F36" ] }, { "n": 3, "para": 1, "text": "Ai2 released the first version on 18 August 2023 as the largest open dataset to date, with a data sheet; the paper by Luca Soldaini and 35 others appeared on arXiv in January 2024 and at ACL 2024, where it won the Best Resource Paper Award.", "claims": [ "F17", "F42", "F53", "F51", "F52" ] }, { "n": 4, "para": 1, "text": "Six versions followed, from v1 in August 2023 to v1.7 in April 2024, which trained OLMo 7B-v1.7 on 1.715 trillion sampled tokens out of 2,532 million documents.", "claims": [ "F4", "F2", "F5", "F12", "F34" ] }, { "n": 5, "para": 2, "text": "Its sources are Common Crawl snapshots of 2020 to 2023 and C4, peS2o's scientific papers, The Stack's code, Wikipedia, Reddit and Project Gutenberg, with RefinedWeb, StarCoder and others added in v1.7.", "claims": [ "F24", "F29", "F30", "F7", "F6" ] }, { "n": 6, "para": 2, "text": "It is English only, a choice its makers say reinforces English as the default language and which removes 61.7% of the web data.", "claims": [ "F23", "F49" ] }, { "n": 7, "para": 3, "text": "Its makers filtered for quality with heuristics after Gopher and Falcon, deduplicated by URL, document and paragraph with a Bloom filter, masked personal information, removed toxic content with fastText classifiers and decontaminated against their evaluation suite, and they call the filters inherently ideological.", "claims": [ "F22", "F25", "F45", "F26", "F27", "F47", "F15", "F28", "F46", "F31", "F40" ] }, { "n": 8, "para": 4, "text": "The release answers a field in which commercial models rarely detail their data, and documents the corpus with its toolkit open-sourced so that others can rebuild, study and critique it, including for the attribution of model output back to pretraining data.", "claims": [ "F37", "F19", "F20", "F39", "F41", "F44", "F35" ] }, { "n": 9, "para": 4, "text": "First released under Ai2's ImpACT license as a medium-risk artifact, it has been under ODC-BY since 15 April 2024, its users also bound by the terms of the original sources.", "claims": [ "F32", "F3", "F16", "F10" ] }, { "n": 10, "para": 4, "text": "A form lets a person request the removal of documents containing personal information about them.", "claims": [ "F21", "F33", "F9", "F48" ] }, { "n": 11, "para": 5, "text": "The card gives each source's provenance, cut-off date and processing, and lists 241 models trained or fine-tuned on the dataset, OLMo among them.", "claims": [ "F8", "F11", "F50" ] }, { "n": 12, "para": 5, "text": "Wikipedia has no article on it; it appears in the article on Ai2 as the training dataset opened with OLMo in February 2024, and Wikipedia's Dolma is a dish.", "claims": [ "F55", "F56", "F57", "F58" ] } ] }, "L_B": { "words": 381, "text": "Dolma is a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books and encyclopedic materials, described in its paper as a three-trillion-token English corpus for research on language model pretraining. It is two things, the dataset and the Dolma Toolkit, and was created as a training corpus for OLMo, a language model from the Allen Institute for AI, which named it Data to feed OLMo's Appetite. Ai2 released the first version on 18 August 2023 as the largest open dataset to date, with a data sheet; the paper by Luca Soldaini and 35 others appeared on arXiv in January 2024 and at ACL 2024, where it won the Best Resource Paper Award. Six versions followed, from v1 in August 2023 to v1.7 in April 2024, which trained OLMo 7B-v1.7 on 1.715 trillion sampled tokens out of 2,532 million documents. Its sources are Common Crawl snapshots of 2020 to 2023 and C4, peS2o's scientific papers, The Stack's code, Wikipedia, Reddit and Project Gutenberg, with RefinedWeb, StarCoder and others added in v1.7. It is English only, a choice its makers say reinforces English as the default language and which removes 61.7% of the web data. Its makers filtered for quality with heuristics after Gopher and Falcon, deduplicated by URL, document and paragraph with a Bloom filter, masked personal information, removed toxic content with fastText classifiers and decontaminated against their evaluation suite, and they call the filters inherently ideological. The release answers a field in which commercial models rarely detail their data, and documents the corpus with its toolkit open-sourced so that others can rebuild, study and critique it, including for the attribution of model output back to pretraining data. First released under Ai2's ImpACT license as a medium-risk artifact, it has been under ODC-BY since 15 April 2024, its users also bound by the terms of the original sources. A form lets a person request the removal of documents containing personal information about them. The card gives each source's provenance, cut-off date and processing, and lists 241 models trained or fine-tuned on the dataset, OLMo among them. Wikipedia has no article on it; it appears in the article on Ai2 as the training dataset opened with OLMo in February 2024, and Wikipedia's Dolma is a dish.", "note": "L_F(B), the field composed in its own ontology (§0.4), realized as KO_B; its compression is P_B" }, "L_BA": { "words": 2002, "text": "Dolma is an open English corpus of three trillion tokens for research on language model pretraining, drawn from web content, scientific papers, code, public-domain books, social media and encyclopedic materials, and released with its documentation and an open-source toolkit for curating such data. The name covers two things, the dataset and the Dolma Toolkit that curated it; on its first release it was announced as the largest open dataset to date, it is catalogued as allenai/dolma and under the title of its paper, and the corpus is itself pretraining data. The Allen Institute for AI (Ai2), building its open language model OLMo since March 2023, created Dolma as OLMo's training corpus and named it accordingly: Dolma stands for 'Data to feed OLMo's Appetite'. The first version was released on 18 August 2023; the paper that describes the corpus, by Luca Soldaini and 35 other authors, appeared on arXiv on 31 January 2024, was revised on 6 June 2024 and was published in the long papers of ACL 2024 (Bangkok, pages 15725–15788), where it received the Best Resource Paper Award. The layer census, which read the paper's abstract after the dataset card returned a rate limit, records the operator as the paper's 36 authors with no institution named in its saved sources and dates the resource from 2024; the dataset card, the toolkit's pages and Ai2's announcement name Ai2 as maker and date the first version to August 2023. Wikipedia has no article on the corpus: its 'Dolma' is a family of stuffed vegetable dishes, with people and a 2012 Tibetan film among the other senses of the name, and the corpus appears in the article on Ai2 (last edited 13 October 2026) as the training dataset whose contents were opened with OLMo in February 2024, which that article calls 'the most open state-of-the-art model available'. The dataset card lists six versions: v1 of August 2023; v1.5 of October 2023, roughly three trillion tokens, used to train OLMo-1B; v1.6 and a sample of roughly ten billion tokens, of January 2024; and v1.7 of April 2024, the default, with new sources, more quality filtering and fuzzy deduplication, used to train OLMo 7B-v1.7. Ai2's figure for the first release gives 5,245 million documents, 5,334 GB and 3,084 billion tokens; the card's table for v1.6 gives 4,367 million documents and 3,059 billion Llama tokens, and its table for v1.7 gives 2,532 million documents, of which sampling yields the 1.715 trillion tokens actually used in training. The census counts the corpus in tokens, three trillion as of the paper's second version, records no unit of record, and leaves later versions and the dataset's current state unread. Its web text comes from 24 Common Crawl snapshots of May 2020 to June 2023, its main content extracted with the CCNet pipeline, and from C4, drawn from an April 2019 snapshot; its scientific papers come from peS2o, 38 million permissibly licensed manuscripts of Semantic Scholar's corpus; its code from The Stack, permissively licensed GitHub repositories; its encyclopedic text from the English and Simple English Wikipedia; and its books from Project Gutenberg, filtered to books primarily in English. In the card's table for v1.6 Project Gutenberg supplies the books, some 56,000 documents and six billion tokens; v1.7 adds RefinedWeb, StarCoder, arXiv and StackExchange by way of RedPajama, Flan, CC News, OpenWebMath, Algebraic Stack and MegaWika to the earlier sources. Corpora of this kind, in the archive's account, are built from the documents available on the crawlable surface, which is the population retrieval systems index (#753). Its makers worked by four principles: follow existing practice, trust the evaluation suite on interventions it can measure, favour decisions that help Ai2's core research directions, and take a harms-based approach to risk. The corpus is English only: a language identifier keeps documents with a likelihood above 50% of being English, a filter that removes 61.7% of the web data, and the makers recognize that the decision 'reinforces the assumption of English as the \"default\" language'. For quality its makers chose heuristics and regular expressions after Gopher and Falcon, setting aside model-based filters that select text resembling Wikipedia pages or books because their biases are hard to detect, and a reviewer of the paper reports that it relies solely on heuristic filtering. The archive's account of corpus filtering places Dolma with C4 and RefinedWeb among corpora whose perplexity-based or classifier-based filters 'keep what looks like Wikipedia; discard the rest' (#98); Dolma's makers do use fastText classifiers to remove harmful or obscene content and to identify language, and they write that 'quality filter' misdescribes filters that 'select text according to criteria that are inherently ideological'. Duplicates are removed by URL, by document and by paragraph with a Bloom filter, which the toolkit provides alongside taggers after Gopher, C4 and OpenWebText; email addresses, phone numbers and IP addresses are masked, a document with more than five such items is dropped, and paragraphs that also occur in the evaluation suite are removed, less than 0.001% of the data by characters. The paper reports analyses and experiments on intermediate states of the corpus to share what its makers learned about curation, and the census records the method as the curation of a mixture of sources with the toolkit released. Such a pipeline, in the archive's account, sheds provenance at each stage: scraping, boilerplate removal, format conversion, deduplication, quality filtering and tokenization strip the signals of the publication layer, the author tags and copyright of an HTML head are almost always discarded by the extractors that prepare text for a training corpus, and 'The body text is what survives' (#127). The paper begins from the scarcity it answers: commercial models rarely detail their data, and open models are often released without their training data or the recipes to reproduce them. Ai2 set openness first among its goals, a dataset that lets other researchers rebuild it, study the relation between the data and any model trained on it, report what they find and critique the curation, and it counted among the uses of open data 'attribution of model output back to pretraining data'. The paper documents the corpus's design principles, construction and contents, a data sheet accompanied the first release, and the toolkit is open source; a reviewer calls Dolma and OLMo 'completely transparent and open-source', and the makers set Dolma apart from other open datasets by its size and its licence. The census reads documentation and toolkit as what can be read back from the corpus, records no search or membership test for a single work, reaches the dataset through Hugging Face and the linked code, and marks it as no closed commercial corpus. The archive counts datasheets for datasets, standardized documentation of training-data provenance, as a step toward accounting for the semantic labor that training data draws on (#623) and as the norm of dataset-level citation and provenance disclosure (#1075), and holds that a training corpus without provenance 'is not a commons' (#727). Dolma was first released under Ai2's ImpACT license as a medium-risk artifact, which required researchers to give contact information and intended uses, to disclose derivatives and distribute them under the same restrictions, and to forgo prohibited uses; on 15 April 2024 the licence changed to ODC-BY. Under ODC-BY the card adds that users are 'also bound any license agreements and terms of use of the original data sources'; the census, without the card, records no licence. Its makers weighed how far web content can be traced back to a real-world individual and created a mechanism, a form named on the card and by a reviewer, through which a person can ask for the removal of documents containing personal information about them. The card records the origin of its documents by source: its table for v1.7 gives each source's provenance, such as Common Crawl by way of an earlier Dolma version, with a cut-off date and the processing applied. Whether a document in Dolma keeps its author's name, its source or its attribution metadata is not recorded in the census's saved sources, and the census lists it, with whether a named work is in Dolma at all and which other makers' models were trained on it, among the questions the resource leaves unanswered; it marks the corpus as no record of authorship and no search tool for single works. The census reads the corpus as settling, at its own scale, which kinds of writing enter the layer and as making its construction inspectable, and leaves outside the abstract what a model makes from a given work and whether the work is carried 'with the metadata that lets it be found and attributed', the third criterion of training-layer literature in the archive's field survey (#186). In the archive's account of training pipelines, an author's style passes into a model's distribution in proportion to its volume and 'no specific author's style survives as a discrete signal that can be elicited and verified', while a membership inference attack can show after training whether a trap sentence placed in a document was in the training set, exact repeated traps being vulnerable to deduplication (#127). The archive names what this shedding does to the writing of the dead: 'the dead can converse at last — and they converse without their names', the tokenizer receiving the body after scraping, extraction, deduplication and packing have detached its source identity (#1070). Its measure of attribution loss, the Provenance Erasure Rate, cannot be computed directly for a model without retrieval, whose source is the entire training corpus (#716); and the archive records an audit (Longpre et al., 2024) of C4, RefinedWeb and Dolma that documents the collapse of web-consent infrastructure (#98). Dolma trained OLMo, a set of open language models, and the dataset card lists 241 models trained or fine-tuned on it, among them models of other makers. A study the archive records examined OLMo's accessible training materials and found that input prevalence alone does not explain the concentration of recurring names in the stories models write (#1695). In the archive's formalization of training-layer literature, the institutional training loop, in which a work is read by a model whose output enters the next corpus and the next model, is the especially important case of recursive reception, and the Provenance Erasure Rate asks what survived of a work where generative uptake asks what became possible because of it (#1543). The archive's field survey holds that the works it surveys are, as a matter of fact, in the corpora, the distinction lying in whether a practice is organized around that fact, and that a literature written to be trained on enters corpora 'themselves degrading under synthetic flood', against which provenance-anchored, human-authored, declared work is the counter-signal (#186). The layer census admits a resource that gathers and holds many other works or records and says what it holds, asks how literature is addressed to a receiving apparatus, what that apparatus makes from it and how transmission, transformation and provenance can be inspected across later reception, and is anchored on #186 and #1543. It places Dolma in its area of corpora, the collections, crawls and datasets whose texts enter machine training, as infrastructure and as a dataset whose question is which kinds of writing a documented pretraining corpus holds and how it was built; by its record the corpus answers which kinds of source a model trained on it has read, its size and language and its construction, and the dataset card, the full paper and the code remain unchecked. Asked 'Dolma dataset' on 1 November 2026, Google's AI Overview composed the corpus accurately and more finely than the census's abstract, from the dataset's own pages and its paper, with a named set for each kind of source and Ai2 as maker, and named no author (Capture Registry). The registry records the version 1.7 figures and the licences as on no source the census saved, the purpose narrowed from research to one model family, and nothing in the composition on what of a book's author or source survives into the corpus.", "rail": [], "note": "L(B ∪ A) in the equal-terms case (§0.4), realized as the knowledge object (objects.KO); its compression is objects.P" } }, "field": [ { "id": "B1", "card": "Hugging Face — allenai/dolma · Datasets at Hugging Face", "fetched": "card and own page; B1.html sha256 ba6f8001031a5e0d, HTTP 200 at 2026-11-01T15:13:22Z; the census's copy (dolma--card) HTTP 429 at 2026-11-01T04:07:51Z, unread", "origin": "card", "pool": "B_D" }, { "id": "B2", "card": "GitHub Pages documentation — dolma | Data and tools for generating and inspecting OLMo pre-training data.", "fetched": "card; B2.html sha256 9b7b9191b66c1ae1, HTTP 200 at 2026-11-01T15:13:22Z", "origin": "card", "pool": "B_D" }, { "id": "B3", "card": "Allen Institute for Artificial Intelligence — Ai2 Dolma: 3 trillion token open corpus for language model pretraining | Ai2", "fetched": "card; B3.html sha256 3ca32b1bc011de1c, HTTP 200 at 2026-11-01T15:13:23Z", "origin": "card", "pool": "B_D" }, { "id": "B4", "card": "arXiv — [2402.00159] Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", "fetched": "card and own page; B4.html sha256 1203d77da94d4612, HTTP 200 at 2026-11-01T15:13:23Z; byte-identical to the census's saved copy (dolma--paper)", "origin": "card", "pool": "B_D" }, { "id": "B5", "card": "Deep (Learning) Focus — Dolma, OLMo, and the Future of Open-Source LLMs", "fetched": "card; B5.html sha256 e46900187043f428, HTTP 200 at 2026-11-01T15:13:24Z", "origin": "card", "pool": "B_D" }, { "id": "B6", "card": "ACL Anthology — Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - ACL Anthology", "fetched": "card; B6.html sha256 08bbcf3ad027b9f6, HTTP 200 at 2026-11-01T15:13:24Z", "origin": "card", "pool": "B_D" }, { "id": "B7", "card": "Wikipedia — Allen Institute for AI (the passage on Dolma)", "fetched": "wikipedia; B7.html sha256 b7669180cb210683, HTTP 200 at 2026-11-01T15:15:19Z; wikitext HTTP 200 (sha256 87b14f4b9a1f79c8); no article for Dolma (Dolma_(dataset) 404, Dolma_(corpus) 429 twice)", "origin": "wikipedia", "pool": "B_R" }, { "id": "B8", "card": "Wikipedia — Dolma (disambiguation) (the record of absence)", "fetched": "wikipedia; B8.wiki sha256 68d67621337c934a, HTTP 200 at 2026-11-01T15:14:03Z; the record of the name's senses on Wikipedia, none the corpus", "origin": "wikipedia", "pool": "B_R" } ], "field_claims": { "F1": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "opening", "quote": "Dolma is a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials.", "claim": "the dataset card: a dataset of 3 trillion tokens from web content, academic publications, code, books and encyclopedic materials", "modality": "stipulation", "kind": "definition", "kernel": true, "const": true, "lineage": "what it is: an open English pretraining corpus", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_def", "also_origin": "own_page" }, "F2": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "news (2024-04-17)", "quote": "2024-04-17: Dolma v1.7 Release. We have released an updated version of Dolma that we used to train our latest OLMo 7B-v1.7 model.", "claim": "v1.7 released 17 April 2024, used to train OLMo 7B-v1.7", "modality": "documented", "kind": "genealogy", "kernel": false, "const": false, "lineage": "versions and size", "date": "2024-04-17", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_v17", "also_origin": "own_page" }, "F3": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "news (2024-04-15)", "quote": "2024-04-15: License Change. We have updated the license of Dolma to ODC-BY .", "claim": "the licence changed to ODC-BY on 15 April 2024", "modality": "documented", "kind": "genealogy", "kernel": true, "const": false, "lineage": "licence and access", "date": "2024-04-15", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_lic_change", "also_origin": "own_page" }, "F4": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Versions (table)", "quote": "At the moment, there are six versions of Dolma available: … v1_7 … 2024-04-15 … Used to train OLMo-7B-v1.7 . New sources, more quality filtering, fuzzy deduplication. … v1_6 … 2024-01-31 … v1_6-sample … A smaller sample of Dolma, with roughly 10 billion tokens. … v1_5 … 2023-10-31 … Used to train OLMo-1B . Roughly 3 trillion tokens. … v1 … 2023-08-18 … The first version of Dolma.", "claim": "six versions: v1 (2023-08-18), v1.5 (2023-10-31, about three trillion tokens, OLMo-1B), v1.6 and a ten-billion-token sample (2024-01-31), v1.7 (2024-04-15, default, new sources, more filtering, fuzzy deduplication)", "modality": "documented", "kind": "genealogy", "kernel": true, "const": false, "lineage": "versions and size", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_versions", "also_origin": "own_page" }, "F5": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Summary Statistics (v1.7), Total", "quote": "Total … 2532.0 … (A subset of total data was used for training of OLMo 7B-v1.7. The token counts are based on the full dataset, whereas taking into account sampling proportion gives the final actual token counts used for training --- 1.715 trillion tokens.)", "claim": "v1.7: 2,532 million documents in total; with sampling, 1.715 trillion tokens actually used in training", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "versions and size", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_v17_total", "also_origin": "own_page" }, "F6": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Summary Statistics (v1.7), Source and Provenance columns", "quote": "Refined Web … StarCoder … Reddit … PushShift API … arXiv … RedPajama v1 … StackExchange … Flan … CC News … OpenWebMath … Algebraic Stack … MegaWika … Wikipedia & Wikibooks", "claim": "the v1.7 sources: Dolma's CC, RefinedWeb, StarCoder, C4, Reddit (PushShift), peS2o, arXiv and StackExchange (RedPajama v1), Flan, CC News, OpenWebMath, Algebraic Stack, Project Gutenberg, MegaWika, Wikipedia and Wikibooks", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "what it holds: sources", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_v17_sources", "also_origin": "own_page" }, "F7": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Summary Statistics (v1.6), Project Gutenberg row", "quote": "Project Gutenberg … books … 20.4 … 0.056 … 4.0 … 6.0", "claim": "v1.6: Project Gutenberg supplies the books, 0.056 million documents, 6.0 billion Llama tokens", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "what it holds: sources", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_pg", "also_origin": "own_page" }, "F8": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Summary Statistics (v1.7), header and first row", "quote": "Source … Provenance … New? … Documents (millions) … OLMo tokens (billions) … Sample Proportion … Cutoff Date … Processing … Dolma's CC … Common Crawl via Dolma v1.6", "claim": "the v1.7 table records, per source, its provenance (the upstream collection it came through), cut-off date and processing", "modality": "documented", "kind": "fact", "kernel": true, "const": true, "lineage": "what of a work's author and source it keeps", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_prov", "also_origin": "own_page" }, "F9": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "opening", "quote": "Want to request removal of personal data? Use this form to notify us of documents containing PII about a specific user.", "claim": "a form to request removal of documents containing personal information about a specific person", "modality": "documented", "kind": "function", "kernel": false, "const": false, "lineage": "personal data and removal", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_pii", "also_origin": "own_page" }, "F10": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Licensing Information", "quote": "We are releasing this dataset under the terms of ODC-BY . By using this dataset, you are also bound any license agreements and terms of use of the original data sources.", "claim": "released under ODC-BY; users also bound by the licences and terms of the original data sources", "modality": "documented", "kind": "fact", "kernel": true, "const": true, "lineage": "licence and access", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_lic", "also_origin": "own_page" }, "F11": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "models trained on the dataset", "quote": "Models trained or fine-tuned on allenai/dolma … Browse 241 models trained on this dataset", "claim": "241 models trained or fine-tuned on the dataset, among them models of other makers (dmis-lab/UltraBERT)", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "models trained on it, and the training loop", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_models", "also_origin": "own_page" }, "F12": { "source": "B1", "origin": "card", "pool": "B_D", "locus": "Summary Statistics (v1.6), Total", "quote": "Total … 11,519 … 4,367 … 2,318 … 3,059", "claim": "v1.6: 11,519 GB, 4,367 million documents, 2,318 billion words, 3,059 billion Llama tokens", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "versions and size", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B1.txt (sha256 bd31e420de096396…) or the card's title and snippet, whitespace-normalised", "key": "hf_v16_total", "also_origin": "own_page" }, "F13": { "source": "B2", "origin": "card", "pool": "B_D", "locus": "opening", "quote": "Dolma is two things: … Dolma Dataset : an open dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials. … Dolma Toolkit : a high-performance toolkit for curating datasets for language modeling", "claim": "Dolma is two things: the dataset and the Dolma Toolkit, a toolkit for curating datasets for language modeling", "modality": "stipulation", "kind": "definition", "kernel": true, "const": true, "lineage": "what it is: an open English pretraining corpus", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B2.txt (sha256 1ceed892a982dccd…) or the card's title and snippet, whitespace-normalised", "key": "gh_two" }, "F14": { "source": "B2", "origin": "card", "pool": "B_D", "locus": "Dolma Dataset", "quote": "It was created as a training corpus for OLMo , a language model from the Allen Institute for AI (AI2).", "claim": "created as a training corpus for OLMo, a language model from the Allen Institute for AI (AI2)", "modality": "documented", "kind": "genealogy", "kernel": true, "const": true, "lineage": "maker: Ai2 and OLMo", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B2.txt (sha256 1ceed892a982dccd…) or the card's title and snippet, whitespace-normalised", "key": "gh_olmo" }, "F15": { "source": "B2", "origin": "card", "pool": "B_D", "locus": "Dolma Toolkit", "quote": "High Performance ⚡: Can process billions of documents concurrently thanks to built-in parallelism. … Built-In Taggers 🏷: Includes ready-to-use taggers commonly used to curate datasets such as Gopher , C4 , and OpenWebText . … Fast Deduplication 🗑: Speedy document deduplication using a Rust Bloom filter.", "claim": "the toolkit: parallel processing of billions of documents, built-in taggers after Gopher, C4 and OpenWebText, document deduplication with a Rust Bloom filter", "modality": "stipulation", "kind": "function", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B2.txt (sha256 1ceed892a982dccd…) or the card's title and snippet, whitespace-normalised", "key": "gh_toolkit" }, "F16": { "source": "B2", "origin": "card", "pool": "B_D", "locus": "Dolma Dataset", "quote": "Dolma is licensed under ODC-BY ; see our blog post for explanation.", "claim": "licensed under ODC-BY", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "licence and access", "date": "undated (the page as fetched 2026-11-01)", "quote_ok": true, "checked_against": "saved text B2.txt (sha256 1ceed892a982dccd…) or the card's title and snippet, whitespace-normalised", "key": "gh_lic" }, "F17": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "opening", "quote": "Today, we release our first data artifact in this project — Dolma ¹, a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials. Openly available for download on the HuggingFace Hub , Dolma is the largest open dataset to date. … For more details, we also release a data sheet (Gebru et al, 2018) as initial documentation.", "claim": "released 18 August 2023 as the project's first data artifact, the largest open dataset to date, with a data sheet", "modality": "self-description", "kind": "genealogy", "kernel": true, "const": false, "lineage": "what it is: an open English pretraining corpus", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_release" }, "F18": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "opening", "quote": "Since March, we at the Allen Institute for AI have been creating OLMo , an open language model to promote the study of large-scale NLP systems.", "claim": "Ai2 has been building OLMo, an open language model, since March 2023", "modality": "documented", "kind": "genealogy", "kernel": false, "const": false, "lineage": "maker: Ai2 and OLMo", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_olmo" }, "F19": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "1. Project Goals, Openness", "quote": "Openness. Lack of access to pretraining corpora alongside corresponding language models has been a major obstacle for the broader research community. We want to create a dataset that gives other researchers the opportunity to independently create better versions of this dataset, study the relationship between the data and any model trained on it, report any issues they observe when inspecting our data, and critique our curation practices", "claim": "openness first among the goals: a dataset others can rebuild, study against models, inspect and critique", "modality": "stipulation", "kind": "function", "kernel": true, "const": false, "lineage": "openness, documentation and the toolkit", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_open" }, "F20": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "1. Project Goals, Openness", "quote": "Beyond transparency, open data is critical for research that has become increasingly important with the proliferation of generative models, such as attribution of model output back to pretraining data.", "claim": "open data is critical for research such as attribution of model output back to pretraining data", "modality": "interpretation", "kind": "interpretation", "kernel": false, "const": false, "lineage": "what of a work's author and source it keeps", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_attrib" }, "F21": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "1. Project Goals, Risk Mitigation", "quote": "Risk Mitigation. We should minimize risk posed to individuals while meeting reproducibility and representativeness requirements. For example, we are concerned to what extent content in web crawled data can be traced back to a real-world individual.", "claim": "risk to individuals minimized; concern over how far web content can be traced back to a real-world individual", "modality": "stipulation", "kind": "function", "kernel": false, "const": false, "lineage": "personal data and removal", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_risk" }, "F22": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "2. Dataset Design Principles", "quote": "Follow existing practices. … Trust the evaluation suite on interventions it can measure; avoid dramatic shifts when it can't. … Favor decisions that help us (Ai2) with our core research directions. … Take a harms-based approach to risk mitigation.", "claim": "four design principles: follow existing practice; trust the evaluation suite where it can measure; favour decisions that help Ai2's core research directions; a harms-based approach to risk", "modality": "stipulation", "kind": "function", "kernel": false, "const": true, "lineage": "curation: filtering", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_favor" }, "F23": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3. How Did We Create Dolma?, English only", "quote": "English only. … We use a fairly permissive threshold, keeping documents that have a likelihood over 50% of being in English. … We recognize that our decision reinforces the assumption of English as the \"default\" language", "claim": "English only: documents kept above a 50% likelihood of English; the makers recognize that the decision reinforces English as the \"default\" language", "modality": "self-description", "kind": "fact", "kernel": true, "const": true, "lineage": "English only", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_english" }, "F24": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Web data", "quote": "Our dataset is derived from 24 Common Crawl snapshots collected between 2020–05 to 2023–06. We use the CCNet pipeline to obtain the main content of each web page in plain text form. Further, we also use the C4 dataset, which is obtained from a Common Crawl snapshot collected in April 2019.", "claim": "web data from 24 Common Crawl snapshots (May 2020 to June 2023), main content extracted with the CCNet pipeline; C4 from an April 2019 snapshot", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "what it holds: sources", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_web" }, "F25": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Quality filtering", "quote": "For example, unigram language models or linear classifiers could be used to select content that broadly resembles Wikipedia pages or books⁴. However, these model-based approaches often have biases that are hard to detect. Following Gopher and Falcon , we opt to use simple heuristics and regular expressions to filter paragraphs.", "claim": "quality filtering by heuristics and regular expressions after Gopher and Falcon, over model-based filters selecting text that resembles Wikipedia or books, whose biases are hard to detect", "modality": "stipulation", "kind": "function", "kernel": true, "const": true, "lineage": "curation: filtering", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_quality" }, "F26": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "Notes (note 3)", "quote": "The term \"quality filter\", while widely used in literature, does not appropriately describe the outcome of filtering a dataset. … the filters used in Dolma and other language models efforts select text according to criteria that are inherently ideological .", "claim": "the filters used in Dolma select text by criteria that are inherently ideological; \"quality filter\" misdescribes them", "modality": "self-description", "kind": "interpretation", "kernel": false, "const": false, "lineage": "curation: filtering", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_ideology" }, "F27": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Deduplication", "quote": "First, in common crawl data, we deduplicate pages based on their URL, keeping only one copy of each. Then, we remove duplicate paragraphs within single documents. Both stages use a Bloom filter data structure.", "claim": "deduplication of web pages by URL and of paragraphs within documents, with a Bloom filter", "modality": "documented", "kind": "function", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_dedup" }, "F28": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Risk mitigation", "quote": "We detect and mask email addresses, phone numbers, and IP addresses. We also remove content that is detected to be harmful or obscene by fasttext classifiers trained on the Jigsaw dataset .", "claim": "email addresses, phone numbers and IP addresses masked; harmful or obscene content removed by fastText classifiers", "modality": "documented", "kind": "function", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_pii" }, "F29": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Diverse sources", "quote": "For Dolma, we leverage Semantic Scholar's corpus by including papers from peS2o , a subset of 38M permissibly licensed scientific manuscripts. We also include Wikipedia and Project Gutenberg. … For Wikipedia, we use the English and Simplified English subset. For the books in Project Gutenberg, we filter for books that are primarily in English.", "claim": "papers from peS2o (38M permissibly licensed manuscripts of Semantic Scholar); English and Simple English Wikipedia; Project Gutenberg books primarily in English", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "what it holds: sources", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_sources" }, "F30": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Just enough code", "quote": "We derive the code subset of Dolma from The Stack , a collection of permissively-licensed GitHub repositories.", "claim": "code from The Stack, permissively licensed GitHub repositories", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "what it holds: sources", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_code" }, "F31": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "3., Decontamination", "quote": "Our decontamination step removes less than 0.001% of the training data by characters, and impacts fewer than 0.02% of documents.", "claim": "decontamination against the evaluation suite removes under 0.001% of the data by characters", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_decon" }, "F32": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "6. Releasing Dolma", "quote": "Dolma is released under Ai2's ImpACT license as a medium-risk artifact. Under this license, researchers must: … Provide their contact information and state their intended use case(s) for accessing Dolma; … Disclose the creation of any derivative based on Dolma; … Distribute derivatives under the same restrictions as the ImpACT license; … Agree not to leverage Dolma in a range of prohibited uses, such as military surveillance or generating disinformation.", "claim": "released under Ai2's ImpACT licence as a medium-risk artifact: contact and intended use given, derivatives disclosed and restricted, prohibited uses forgone", "modality": "documented", "kind": "fact", "kernel": true, "const": false, "lineage": "licence and access", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_impact" }, "F33": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "6. Releasing Dolma", "quote": "Finally, as part of our risk mitigation strategy, we created a mechanism to allow the removal of personal data upon request.", "claim": "a mechanism for removing personal data on request", "modality": "documented", "kind": "function", "kernel": true, "const": false, "lineage": "personal data and removal", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_removal" }, "F34": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "overview figure (image alt text, in B3.html)", "quote": "Overview of Dolma by subset type. Sources include CommonCrawl, C4, peS2o, The Stack, Project Gutenberg, and Wikipedia with a total of 5,245 millions of documents at 5,334 GB, and 3,084 billions of tokens.", "claim": "the first release: 5,245 million documents, 5,334 GB, 3,084 billion tokens, from Common Crawl, C4, peS2o, The Stack, Project Gutenberg and Wikipedia", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "versions and size", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_alt" }, "F35": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "5. How Does Dolma Compare With Other Open Datasets?", "quote": "Dolma differentiates itself from other datasets on two key aspects. First, it is significantly larger than other open datasets. Second, it is released under Ai2's impact license, which was designed to balance ease of access with mitigation of potential risk in distributing large datasets.", "claim": "larger than other open datasets, and released under a licence balancing access with risk", "modality": "self-description", "kind": "interpretation", "kernel": false, "const": false, "lineage": "openness, documentation and the toolkit", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_differ" }, "F36": { "source": "B3", "origin": "card", "pool": "B_D", "locus": "Notes", "quote": "Fun fact: Dolma stands for \" Data to feed OLMo's Appetite \"", "claim": "the name: Data to feed OLMo's Appetite", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "maker: Ai2 and OLMo", "date": "2023-08-18 (the announcement's date)", "quote_ok": true, "checked_against": "saved text B3.txt (sha256 006aa43afa5bb834…), the raw B3.html (for the figure's alt text) or the card's snippet, whitespace-normalised", "key": "ai_name" }, "F37": { "source": "B4", "origin": "card", "pool": "B_D", "locus": "Abstract", "quote": "Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them.", "claim": "pretraining corpora are seldom discussed; commercial models rarely detail their data; open models often come without training data or recipes", "modality": "interpretation", "kind": "interpretation", "kernel": false, "const": false, "lineage": "openness, documentation and the toolkit", "date": "2024-01-31 (arXiv v1; v2 2024-06-06)", "quote_ok": true, "checked_against": "saved text B4.txt (sha256 81726256243b7fe6…) or the card's title and snippet, whitespace-normalised", "key": "ax_context", "also_origin": "own_page" }, "F38": { "source": "B4", "origin": "card", "pool": "B_D", "locus": "Abstract", "quote": "To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials.", "claim": "for research on language model pretraining, a three-trillion-token English corpus from web content, scientific papers, code, public-domain books, social media and encyclopedic materials", "modality": "stipulation", "kind": "definition", "kernel": true, "const": true, "lineage": "what it is: an open English pretraining corpus", "date": "2024-01-31 (arXiv v1; v2 2024-06-06)", "quote_ok": true, "checked_against": "saved text B4.txt (sha256 81726256243b7fe6…) or the card's title and snippet, whitespace-normalised", "key": "ax_def", "also_origin": "own_page" }, "F39": { "source": "B4", "origin": "card", "pool": "B_D", "locus": "Abstract", "quote": "We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents.", "claim": "documented: design principles, construction, contents", "modality": "self-description", "kind": "fact", "kernel": false, "const": false, "lineage": "openness, documentation and the toolkit", "date": "2024-01-31 (arXiv v1; v2 2024-06-06)", "quote_ok": true, "checked_against": "saved text B4.txt (sha256 81726256243b7fe6…) or the card's title and snippet, whitespace-normalised", "key": "ax_doc", "also_origin": "own_page" }, "F40": { "source": "B4", "origin": "card", "pool": "B_D", "locus": "Abstract", "quote": "We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices.", "claim": "analyses and experiments on intermediate states of the corpus, on data curation practice", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "2024-01-31 (arXiv v1; v2 2024-06-06)", "quote_ok": true, "checked_against": "saved text B4.txt (sha256 81726256243b7fe6…) or the card's title and snippet, whitespace-normalised", "key": "ax_analyses", "also_origin": "own_page" }, "F41": { "source": "B4", "origin": "card", "pool": "B_D", "locus": "Abstract", "quote": "Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation.", "claim": "the curation toolkit open-sourced for reproduction and further research", "modality": "documented", "kind": "fact", "kernel": true, "const": false, "lineage": "openness, documentation and the toolkit", "date": "2024-01-31 (arXiv v1; v2 2024-06-06)", "quote_ok": true, "checked_against": "saved text B4.txt (sha256 81726256243b7fe6…) or the card's title and snippet, whitespace-normalised", "key": "ax_toolkit", "also_origin": "own_page" }, "F42": { "source": "B4", "origin": "card", "pool": "B_D", "locus": "title block; Comments; Submission history", "quote": "by Luca Soldaini and 35 other authors … Accepted at ACL 2024; Dataset: this https URL Code: this https URL … [v1] Wed, 31 Jan 2024 20:29:50 UTC (7,715 KB) … [v2] Thu, 6 Jun 2024 18:46:40 UTC (6,680 KB)", "claim": "the paper by Luca Soldaini and 35 others: v1 31 January 2024, v2 6 June 2024, accepted at ACL 2024, linking the dataset and code", "modality": "documented", "kind": "genealogy", "kernel": false, "const": false, "lineage": "the paper: authors, versions, venue, award", "date": "2024-01-31 (arXiv v1; v2 2024-06-06)", "quote_ok": true, "checked_against": "saved text B4.txt (sha256 81726256243b7fe6…) or the card's title and snippet, whitespace-normalised", "key": "ax_meta", "also_origin": "own_page" }, "F43": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "Dolma: an Open Corpus … [1]", "quote": "In [1], authors from AI2 solve this problem by releasing Data for Open Language Models’ Appetite (Dolma), a fully-open pretraining dataset containing over three trillion tokens of English text", "claim": "a reviewer: authors from AI2 solve the problem by releasing a fully-open pretraining dataset of over three trillion tokens of English text", "modality": "attributed", "kind": "interpretation", "kernel": true, "const": false, "lineage": "what it is: an open English pretraining corpus", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_def" }, "F44": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "This overview", "quote": "Unlike prior work in open-source LLMs, Dolma and OLMo are completely transparent and open-source. The process of constructing the Dolma pretraining corpus, which is released under the AI2 impact license , is fully-documented in [1].", "claim": "a reviewer: Dolma and OLMo completely transparent and open-source; the corpus under the AI2 impact license, its construction fully documented", "modality": "attributed", "kind": "interpretation", "kernel": false, "const": false, "lineage": "openness, documentation and the toolkit", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_open" }, "F45": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "Quality filtering", "quote": "In [1], authors choose to avoid model-based filtering techniques and rely solely upon heuristic filtering methods.", "claim": "a reviewer: the authors avoid model-based filtering and rely solely on heuristic filtering", "modality": "attributed", "kind": "fact", "kernel": true, "const": false, "lineage": "curation: filtering", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_heur" }, "F46": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "Content filtering", "quote": "To detect PII, a series of regular expression are adopted to find spans of text corresponding to email addresses, IP addresses, and phone numbers. From here, these spans of text are either masked, or the entire document is removed from the corpus (if more than 5 pieces of PII are detected in a single document).", "claim": "a reviewer: PII spans masked, or the whole document removed where more than five are found", "modality": "attributed", "kind": "fact", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_pii" }, "F47": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "Deduplication", "quote": "In [1], authors perform three stages of deduplication for web text: … URL-based : eliminates web pages that are scraped multiple times. … Document-based : eliminates pages that contain exactly the same text. … Paragraph-based : eliminates individual paragraphs with the same text.", "claim": "a reviewer: three stages of web deduplication, by URL, by document and by paragraph", "modality": "attributed", "kind": "fact", "kernel": false, "const": false, "lineage": "curation: deduplication, masking, decontamination", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_dedup" }, "F48": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "Dolma design principles", "quote": "To minimize risk of harm to individuals, one can easily request— using this form —for their data to be removed from the Dolma corpus.", "claim": "a reviewer: anyone can request removal of their data from the corpus", "modality": "attributed", "kind": "fact", "kernel": false, "const": false, "lineage": "personal data and removal", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_removal" }, "F49": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "Language filtering", "quote": "Typically, this language filtering component, which is applied to both web and conversation data in Dolma, eliminates a large amount of source data (e.g., 61.7% of web data is removed).", "claim": "a reviewer: language filtering removes 61.7% of the web data", "modality": "attributed", "kind": "fact", "kernel": false, "const": false, "lineage": "English only", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_lang" }, "F50": { "source": "B5", "origin": "card", "pool": "B_D", "locus": "OLMo", "quote": "OLMo is a set of truly open LLMs that are pretrained on Dolma.", "claim": "a reviewer: OLMo, a set of open LLMs pretrained on Dolma", "modality": "attributed", "kind": "fact", "kernel": false, "const": false, "lineage": "models trained on it, and the training loop", "date": "2024-02-20 (the post's date)", "quote_ok": true, "checked_against": "saved text B5.txt (sha256 ad92d43ad246e79f…) or the card's title and snippet, whitespace-normalised", "key": "dl_olmo" }, "F51": { "source": "B6", "origin": "card", "pool": "B_D", "locus": "record", "quote": "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) … August … Bangkok, Thailand … 15725–15788", "claim": "published in the ACL 2024 proceedings (long papers), Bangkok, August 2024, pp. 15725–15788", "modality": "documented", "kind": "genealogy", "kernel": true, "const": false, "lineage": "the paper: authors, versions, venue, award", "date": "2024-08 (ACL 2024, the proceedings)", "quote_ok": true, "checked_against": "saved text B6.txt (sha256 781aa6ce3c0caa4a…) or the card's title and snippet, whitespace-normalised", "key": "acl_proc" }, "F52": { "source": "B6", "origin": "card", "pool": "B_D", "locus": "record, Award", "quote": "Award: Best Resource Paper Award", "claim": "Best Resource Paper Award", "modality": "documented", "kind": "fact", "kernel": true, "const": false, "lineage": "the paper: authors, versions, venue, award", "date": "2024-08 (ACL 2024, the proceedings)", "quote_ok": true, "checked_against": "saved text B6.txt (sha256 781aa6ce3c0caa4a…) or the card's title and snippet, whitespace-normalised", "key": "acl_award" }, "F53": { "source": "B6", "origin": "card", "pool": "B_D", "locus": "authors", "quote": "Luca Soldaini , … Rodney Kinney , … Jesse Dodge , … Kyle Lo", "claim": "thirty-six authors, Luca Soldaini first and Kyle Lo last", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "the paper: authors, versions, venue, award", "date": "2024-08 (ACL 2024, the proceedings)", "quote_ok": true, "checked_against": "saved text B6.txt (sha256 781aa6ce3c0caa4a…) or the card's title and snippet, whitespace-normalised", "key": "acl_authors" }, "F54": { "source": "B6", "origin": "card", "pool": "B_D", "locus": "Abstract", "quote": "To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus", "claim": "the ACL record's abstract, the arXiv abstract word for word", "modality": "stipulation", "kind": "definition", "kernel": false, "const": false, "lineage": "what it is: an open English pretraining corpus", "date": "2024-08 (ACL 2024, the proceedings)", "quote_ok": true, "checked_against": "saved text B6.txt (sha256 781aa6ce3c0caa4a…) or the card's title and snippet, whitespace-normalised", "key": "acl_abs" }, "F55": { "source": "B7", "origin": "wikipedia", "pool": "B_R", "locus": "Olmo model family", "quote": "In February 2024, 1B and 7B parameter variants of the model were open-sourced, including code, model weights with intermediate snapshots and logs, and contents of their Dolma training dataset (a three-trillion-token English corpus drawn from web text, academic publications, code, books, social media, and encyclopedic sources),", "claim": "Wikipedia (Ai2 article): OLMo open-sourced in February 2024 with the contents of its Dolma training dataset, a three-trillion-token English corpus", "modality": "documented", "kind": "fact", "kernel": true, "const": false, "lineage": "what it is: an open English pretraining corpus", "date": "2026-10-13 (the article's last edit)", "quote_ok": true, "checked_against": "saved text B7.txt (sha256 3083e6a9400a1f4b…) or the wikitext B7.wiki, whitespace-normalised", "key": "wp_passage" }, "F56": { "source": "B7", "origin": "wikipedia", "pool": "B_R", "locus": "Olmo model family", "quote": "making it the most open state-of-the-art model available.", "claim": "Wikipedia (Ai2 article): OLMo the most open state-of-the-art model available", "modality": "attributed", "kind": "interpretation", "kernel": false, "const": false, "lineage": "models trained on it, and the training loop", "date": "2026-10-13 (the article's last edit)", "quote_ok": true, "checked_against": "saved text B7.txt (sha256 3083e6a9400a1f4b…) or the wikitext B7.wiki, whitespace-normalised", "key": "wp_open" }, "F57": { "source": "B7", "origin": "wikipedia", "pool": "B_R", "locus": "footer", "quote": "This page was last edited on 13 October 2026, at 05:08", "claim": "the Ai2 article last edited 13 October 2026", "modality": "documented", "kind": "fact", "kernel": false, "const": false, "lineage": "the name on Wikipedia", "date": "2026-10-13 (the article's last edit)", "quote_ok": true, "checked_against": "saved text B7.txt (sha256 3083e6a9400a1f4b…) or the wikitext B7.wiki, whitespace-normalised", "key": "wp_edit" }, "F58": { "source": "B8", "origin": "wikipedia", "pool": "B_R", "locus": "page", "quote": "'''[[Dolma]]''' is a family of stuffed vegetable dishes. … '''Dolma''' may also refer to: … ==People== … * ''[[Dolma (film)]]'', a 2012 Tibetan film", "claim": "the name on Wikipedia: a family of stuffed vegetable dishes, people, a 2012 Tibetan film; no dataset", "modality": "documented", "kind": "fact", "kernel": true, "const": false, "lineage": "the name on Wikipedia", "date": "2026-11-01 (the page as fetched)", "quote_ok": true, "checked_against": "the wikitext B8.wiki (sha256 68d67621337c934a…), whitespace-normalised", "key": "wd_senses" } }, "field_sources": { "B1": "Hugging Face — allenai/dolma · Datasets at Hugging Face", "B2": "GitHub Pages documentation — dolma | Data and tools for generating and inspecting OLMo pre-training data.", "B3": "Allen Institute for Artificial Intelligence — Ai2 Dolma: 3 trillion token open corpus for language model pretraining | Ai2", "B4": "arXiv — [2402.00159] Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", "B5": "Deep (Learning) Focus — Dolma, OLMo, and the Future of Open-Source LLMs", "B6": "ACL Anthology — Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - ACL Anthology", "B7": "Wikipedia — Allen Institute for AI (the passage on Dolma)", "B8": "Wikipedia — Dolma (disambiguation) (the record of absence)" }, "pools": { "B_D": [ "B1", "B2", "B3", "B4", "B5", "B6" ], "B_R": [ "B7", "B8" ], "note": "v0.10 §0.3, §3.1(a): B_D, every card on the rail; every source T's body names is a card (T's six inline Google redirect links each resolve to a card URL: 'Allen Institute for Artificial Intelligence (Ai2)' to card 3, 'AllenAI Dolma on Hugging Face' to card 1, 'OLMo' to card 5, 'Dolma ACL Research Paper' to card 6, the markers to cards 1, 2 and 4), so no body-named source stands apart. The census's two own pages are cards 1 (dolma--card) and 4 (dolma--paper), carried once in B_D. B_R: the Wikipedia reference sources (B7, the Ai2 article at its Dolma passage; B8, the disambiguation page as the record of absence). The census read no third-party source for the row. T against L_F(B_D) uses B1–B6 alone." }, "delta": { "T_vs_LB": "At 'Dolma dataset' (T1, six cards, B1–B6, the disclosed pool B_D; B7 and B8, the Wikipedia reference sources, stand outside this comparison, §0.3), T composes a definition, an access line, five bulleted features and a closing offer of two questions. Composed from the cards, at their grain or coarser: the definition (F38, F1, F13) with Ai2 as maker and OLMo as purpose (F14, F18, F50); 'transparency' (F19); the sources at the card's resolution, a named set for each kind (F6, F24, F29, F30, F7); the v1.7 training tokens (F2, F5); the pipeline (F25, F27, F47, F28, F46, F31, F15, F41); the licences (F10, F16, F32, F44); the paper by venue (F51, F54). The award (F52) and the reviewer's 'fully-open' (F43) are composed on the rail only, in card snippets (§5.0). Available, not composed: the six versions and their dates (F4); the licence change of 15 April 2024 (F3); the per-source Provenance column (F8); the removal form for personal data (F9, F33, F48) and the concern over tracing web content to individuals (F21); the 241 models trained on the dataset (F11); the first release and its data sheet (F17); the use of open data for 'attribution of model output back to pretraining data' (F20); the design principles (F22); English only as a stated limit (F23, F49); the filters as 'inherently ideological' (F26); the heuristic-only filtering (F45); the paper's context, documentation and analyses (F37, F39, F40); its authors and versions (F42, F53); the name (F36); the claim to size and licence as Dolma's difference (F35); the v1.6 totals (F12). Limit dropped: ODC-BY is composed without the card's qualification that users are 'also bound any license agreements and terms of use of the original data sources' (F10); 'English corpus' without the makers' statement that the choice reinforces English as the default (F23); the curation without the makers' note on its criteria (F26). Substituted reading: (1) the figures of the first release, 5,245 million documents and 5,334 GB (F34, card 3, 2023), composed under version 1.7 ('across ~5.2 billion documents totaling over 5,300 GB'), whose own table gives 2,532 million documents (F5, card 1); (2) the ImpACT licence, superseded on the maker's own card (F3), composed as current beside its successor; (3) toxicity composed as masked ('PII/toxicity masking'), where the makers mask personal information and remove harmful or obscene content (F28). Force raised: the ImpACT licence, a medium-risk licence with conditions on access, derivatives and uses (F32), composed among 'permissive research licenses'; the OLMo purpose (F14) composed as 'Designed and used specifically to pre-train the OLMo … family', against the paper's research purpose (F38) and the 241 models of the card (F11). Composed without attribution, by grain: the maker at the coarse grain of common knowledge (Ai2), recorded; the authors of the paper unnamed. Outside the cards: 'permissive'. The closing offer replaces no content and adds an invitation to code and scripts (recorded under §6.2 as an addition, not as content replaced by use).", "T_vs_LB_reading": "v0.10 §0.5: the first diff is ontological and narrow. T and the field agree on the entity, Ai2's open English pretraining corpus; T's ontology is the product sheet's (what, who, where to get it, scale, sources, pipeline, licence, use), composed at the cards' finer grain and from the card most recent in its facts. What it keeps out of its own cards is the corpus's history in time (versions, the licence change, the first release's figures kept apart from v1.7's) and every statement the makers make of the corpus's limits and of the people in it (English as default, the ideological criteria of filtering, the users bound by the sources' terms, removal on request, tracing back to individuals and attribution back to pretraining data).", "LB_vs_LBA": "Admission adds, in one claim space with the field's: the census's question of what a document in the corpus keeps of its author, source or attribution, unanswered in its sources and in the field, which records provenance per source collection (TLL-12, TLL-21, TLL-31 beside the card's Provenance column); its reading that the corpus settles at its own scale which kinds of writing enter the layer and makes construction inspectable, leaving open whether a work is carried 'with the metadata that lets it be found and attributed' (TLL-22, TLL-23), the third criterion of training-layer literature in #186 (O186-01); #127's account of the pipeline that sheds a work's provenance before the corpus, the HTML author tags discarded by the extractors that prepare training text, style surviving only in proportion to volume, and membership inference as the test of whether a text was in a training set (O127-01 to O127-04), which meets the field's own stages (CCNet extraction, deduplication, filtering, masking) and its 'attribution of model output back to pretraining data'; #1070's dead who 'converse without their names' (O1070-01); #716's limit that PER cannot be computed for a model without retrieval (O716-01); #753's corpora built from the crawlable surface (O753-01), on the field's Common Crawl snapshots; #623 and #1075's datasheets as provenance accounting (O623-01, O1075-01), on the field's data sheet, and #727's corpus without provenance (O727-01); #1543's institutional training loop and its PER/uptake pair (O1543-01, O1543-02), and #186's works in the corpora and the corpora under synthetic flood (O186-02, O186-03), on the field's OLMo and its 241 models; #1695's record that OLMo's training materials do not explain an output concentration by prevalence (D1695-01); and #98's two claims, the consent audit naming Dolma (D98-02) and the placing of Dolma among filters that 'keep what looks like Wikipedia' (D98-01), the latter contested by the field's heuristic filtering and composed beside it (P_coh). The census's own record, read beside the field, differs from it on the maker (no institution, against Ai2 on three cards), the founding (2024, against the 2023 first release), the licence (unrecorded, against ImpACT then ODC-BY), and the models trained on it (unanswered, against the card's list); these are carried as the record's claims with the field's beside them (marks in the plan) and reported. Through the change of ontology the entity keeps the field's definition and gains a sense the field holds only per source: the corpus as a place where a work's author and source either travel with it or are shed.", "alignment": { "T1_cards_claims": [ "F1", "F2", "F3", "F4", "F5", "F6", "F7", "F8", "F9", "F10", "F11", "F12", "F13", "F14", "F15", "F16", "F17", "F18", "F19", "F20", "F21", "F22", "F23", "F24", "F25", "F26", "F27", "F28", "F29", "F30", "F31", "F32", "F33", "F34", "F35", "F36", "F37", "F38", "F39", "F40", "F41", "F42", "F43", "F44", "F45", "F46", "F47", "F48", "F49", "F50", "F51", "F52", "F53", "F54" ], "T1_composed": [ "F38", "F1", "F13", "F14", "F18", "F19", "F2", "F5", "F34", "F6", "F24", "F29", "F30", "F7", "F25", "F27", "F47", "F28", "F46", "F31", "F15", "F41", "F10", "F16", "F32", "F44", "F50", "F51", "F54" ], "T1_composed_rail_only": [ "F52", "F43" ], "T1_composed_note": "F34 composed under v1.7 (substituted reading: the first release's figures); F32 composed as current and 'permissive' (superseded by F3; force raised); F10 composed without its qualification (limit dropped); F28 toxicity composed as masked (substituted reading); F14 composed as 'specifically' (force raised); the sources at card 1's grain", "T1_outside_cards": [ "'permissive' (T1-6: no field claim calls either licence permissive; B3 calls ImpACT a medium-risk licence with conditions)", "'version 1.7 … across ~5.2 billion documents totaling over 5,300 GB' (T1-3: the figures are card 3's for the first release; card 1's v1.7 table gives 2,532 million documents)" ], "rule": "a field claim counts as composed when T states it at its source's grain or coarser, in the body or a card snippet (§5.0, the part recorded); B_D only (B7 and B8 excluded, §0.3); the transcript read for this alignment only" }, "pool_note": "T against L_F(B_D): B1–B6; L_F(B) against L(B ∪ A): B1–B8 with A" }, "kernel": [ { "K": "K1", "claim": "`TLL-12`", "source": "traininglayerliterature.org layer census record", "M_src": "interpretation", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "missing distinction: the field records provenance per source collection (F8) and no field claim says what a document keeps of its author or source; T composes neither", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K2", "claim": "`TLL-21`", "source": "traininglayerliterature.org layer census record", "M_src": "interpretation", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "missing distinction: no field claim offers a membership test for a named work; the field names attribution of output back to pretraining data as a use of open data (F20)", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K3", "claim": "`TLL-22`", "source": "traininglayerliterature.org layer census record", "M_src": "interpretation", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F38, F6 (the kinds and sets of source), read as settling at the corpus's scale which writing enters the layer", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K4", "claim": "`O127-01`", "source": "#127", "M_src": "interpretation", "sense": "how it was made", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F24, F27, F25 (the pipeline's stages), read as the stages that shed provenance", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K5", "claim": "`O127-02`", "source": "#127", "M_src": "interpretation", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F24 (CCNet extracts 'the main content of each web page in plain text form')", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K6", "claim": "`O127-04`", "source": "#127", "M_src": "attributed", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "missing distinction: the field has no membership test; F20 names attribution back to pretraining data as a research use", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K7", "claim": "`O1070-01`", "source": "#1070", "M_src": "interpretation", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F47, F27 (deduplication by URL, document and paragraph)", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K8", "claim": "`D98-01`", "source": "#98", "M_src": "interpretation", "sense": "how it was made", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "rival claim: F25, F45 (heuristic filtering); F28 (fastText classifiers for toxicity) bears on 'classifier-based'; the field claim can win on the same terms (§8.2)", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K9", "claim": "`O753-01`", "source": "#753", "M_src": "interpretation", "sense": "what it holds", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F24 (24 Common Crawl snapshots)", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K10", "claim": "`O623-01`", "source": "#623", "M_src": "interpretation", "sense": "openness, documentation and the toolkit", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F17 (a data sheet with the first release), F39", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K11", "claim": "`O727-01`", "source": "#727", "M_src": "interpretation", "sense": "openness, documentation and the toolkit", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "none: the field has no claim on a corpus as a commons", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K12", "claim": "`O716-01`", "source": "#716", "M_src": "self-description", "sense": "what of a work's author and source it keeps", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "none: the field states no measure of attribution loss", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K13", "claim": "`O1543-01`", "source": "#1543", "M_src": "stipulation", "sense": "models trained on it, and the training loop", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "qualified claim: F50, F11 (OLMo and 241 models trained on the corpus)", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K14", "claim": "`D1695-01`", "source": "#1695", "M_src": "attributed", "sense": "models trained on it, and the training loop", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "none: the field has no claim on what OLMo makes from its training data", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K15", "claim": "`TLL-03`", "source": "traininglayerliterature.org layer census record", "M_src": "documented", "sense": "who made it, and when", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "rival claim: F14, F18 (Ai2 named as maker); the record's 'no institution' holds for its one read source", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" }, { "K": "K16", "claim": "`CAP-1-05`", "source": "Capture Registry", "M_src": "interpretation", "sense": "the composition at the name", "qualifiers carried": "—", "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not (uptake §8.3))", "contrast": "rival claim: T1-5 (the pipeline composed with no statement of what of a document's author or source survives)", "resolver": "a later capture at 'Dolma dataset' (ADDR-5a024e0d48d0) on Google AI Overview, seated in the Capture Registry (§8.3)" } ], "kernel_T": [ { "claim": "F3", "note": "the licence change of 15 April 2024 (card 1), absent; T composes the superseded ImpACT licence beside ODC-BY" }, { "claim": "F10", "note": "users 'also bound any license agreements and terms of use of the original data sources' (card 1), absent" }, { "claim": "F34", "note": "the first release's 5,245 million documents and 5,334 GB (card 3), composed under v1.7" }, { "claim": "F8", "note": "the per-source Provenance column (card 1), absent" }, { "claim": "F23", "note": "English only as reinforcing English as the 'default' language (card 3), absent; T keeps 'English corpus'" }, { "claim": "F26", "note": "filters selecting text by 'inherently ideological' criteria (card 3), absent" }, { "claim": "F9", "note": "the removal form for personal data (card 1; cards 3 and 5), absent" }, { "claim": "F11", "note": "241 models trained on the dataset (card 1), absent; T composes 'specifically' OLMo" }, { "claim": "F42", "note": "the paper's authors (cards 4 and 6), absent; T names no author (CAP-1-02)" } ], "ledger": { "archive": "datasets/negative-of-the-negative/v2/ledgers/dolma/ledger-archive.json", "field": "datasets/negative-of-the-negative/v2/ledgers/dolma/field.json", "selection": "datasets/negative-of-the-negative/v2/traversal/dolma/", "config": "datasets/negative-of-the-negative/v2/panel/configs/dolma.json", "intake": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/", "audit": null }, "ontology": { "composition": "equal", "composition_note": "equal-terms case (v0.9 §0.9, §4.1, unchanged in v0.10): one claim space; the field's and the archive's claims ordered together by genealogy within each sense; the field's claims composed as received, the archive's as claims on the same object", "case": { "value": "equal", "statement": "The archive reads the field's entity as the entity. The layer census describes Dolma from its paper's abstract and keeps the resource's terms (kind, record object, method, holdings, count) apart from its own reading (tll_reading), which asks of the same corpus whether a work is carried 'with the metadata that lets it be found and attributed' (#186) and reads it as settling which kinds of writing enter the layer. #127 and #1070 describe what a training pipeline of this kind sheds of a work's author and source; #753, #623, #1075 and #727 place training corpora in the archive's accounts of retrievability and provenance; #1543 and #186 place the corpus in the training loop; #98 and #1695 name it. None of these reads Dolma as made by reception or contests that it is Ai2's open English pretraining corpus: they add senses (what of a work's author and source survives into the corpus, membership, provenance accounting, the loop) on the field's object. One archive claim contradicts the field (#98 on the filters); both are composed. ρ = extends; the equal-terms case.", "loci": [ { "dep": "TLL", "locus": "census.json rows[slug == 'dolma'].tll_reading", "quote": "A corpus that states its own sources and releases its curation toolkit: it settles, at the scale of the corpus, which kinds of writing enter the layer ('public-domain books', 'web content', 'scientific papers'), and makes the construction itself inspectable." }, { "dep": "TLL", "locus": "census.json rows[slug == 'dolma'].tll_reading", "quote": "What a model makes from a given work, and whether the work is carried with 'the metadata that lets it be found and attributed' (#186), are outside the saved abstract; the dataset card that would show document fields was rate-limited." }, { "dep": 186, "locus": "What is training-layer literature? ¶2, the third criterion (line 24)", "quote": "Third, persistence: the work is placed where the training and retrieval layers actually look, with the metadata that lets it be found and attributed." }, { "dep": 127, "locus": "§IV.2 HTML metadata (line 463)", "quote": "HTML head metadata — author tags, copyright, Open Graph, Twitter cards — is parsed by search engines and social media unfurl bots, but is almost always discarded by readability extractors that prepare text for training corpus inclusion. The training-data pipeline retains the *body text* and the *URL* (sometimes)." }, { "dep": 1543, "locus": "§10 Formal consolidation (line 264)", "quote": "The institutional training loop is the especially important case: x → M_t → y → corpus_{t+1} → M_{t+1} → y′ — T's fixed institutional iteration." }, { "dep": 98, "locus": "Part I, What the Filter Does ¶2 (line 35)", "quote": "C4, RefinedWeb, and Dolma all employ perplexity-based or classifier-based filters that share the same logic: keep what looks like Wikipedia; discard the rest." } ] }, "E_F": { "type": "C (public entity: a documented pretraining corpus and its curation toolkit)", "definition": "Dolma, an open English corpus of three trillion tokens for research on language model pretraining, built by the Allen Institute for AI (Ai2) as the training corpus for OLMo from web content (Common Crawl, C4), scientific papers (peS2o), code (The Stack), public-domain books (Project Gutenberg), social media and encyclopedic materials, filtered by heuristics, deduplicated with Bloom filters, its personal information masked and toxic content removed; first released 18 August 2023 under Ai2's ImpACT licence, in six versions to v1.7 (April 2024), under ODC-BY since 15 April 2024; documented in a paper of Luca Soldaini and 35 others (ACL 2024, Best Resource Paper) and released with the open Dolma Toolkit; English only by design, its provenance recorded per source collection", "claims": [ "F38", "F1", "F13", "F14", "F18", "F24", "F29", "F30", "F25", "F27", "F28", "F17", "F4", "F3", "F42", "F52", "F23", "F8" ] }, "E_A": { "definition": "the same corpus, in the traininglayerliterature.org layer census record and the archive's deposits: pretraining data that settles at its own scale which kinds of writing enter the layer and makes its construction inspectable, infrastructure to the census's question of how literature addressed to a receiving apparatus can be traced through it; whether a document in it keeps its author's name, source or attribution metadata, and whether a named work is in it, are unrecorded; its pipeline is of the kind the archive describes as shedding a work's provenance before the corpus, the body text surviving and the dead conversing 'without their names' (#127, #1070), with membership inference as the test of what was in it (#127); a training corpus without provenance 'is not a commons' (#727), datasheets a step toward the accounting (#623, #1075); its web text comes from the crawlable surface retrieval systems index (#753); it stands in the institutional training loop where PER asks what survived of a work (#1543); and #98 places its filters among those that 'keep what looks like Wikipedia', a claim the field contests", "claims": [ "TLL-11", "TLL-22", "TLL-23", "TLL-12", "TLL-21", "TLL-06", "O186-01", "O127-01", "O127-02", "O127-04", "O1070-01", "O727-01", "O623-01", "O753-01", "O1543-01", "O1543-02", "D98-01" ], "defining_papers": [ 186, 1543, 127, 1070, 727, 716, 753, 623, 1075, 98, 1695 ], "defining_records": [ "traininglayerliterature.org layer census record (census.json rows[slug == 'dolma'], organizing_question, admission_rule, areas, anchor), https://www.traininglayerliterature.org/aggregators/dolma", "Capture Registry entry dolma-dataset-aio-20261101 (OBS-6dd2d86d5018), https://www.alexanarch.org/captures/dolma-dataset-aio-20261101/" ] }, "rho": { "value": "extends", "status": "read from the census record, the capture record and the deposits that bear on the entity and its classes (#186 and #1543, the census anchors; #127, #1070, #727, #716, #753, #623, #1075, #98, #1695), 2026-11-01", "statement": "The archive reads Dolma as its cards present it and adds, on the same object, the question of what a document keeps of its author and source (TLL-12, TLL-21, TLL-22, TLL-23), the pipeline as the place where provenance is shed (#127, #1070), membership as testable (#127), the limit of its own measure (#716), the corpus in the accounts of provenance and retrievability (#623, #1075, #727, #753), and the training loop (#1543, #186, #1695). One archive claim contradicts field claims: #98's perplexity-based or classifier-based filters against the makers' heuristics; both are composed (P_coh).", "claims": [ "TLL-12", "TLL-21", "TLL-22", "TLL-23", "O127-01", "O127-02", "O127-04", "O1070-01", "O716-01", "O623-01", "O1075-01", "O727-01", "O753-01", "O1543-01", "O186-01", "D1695-01", "D98-01" ], "guard": "ρ = extends claims an addition and no more: the archive's claims do not make the field's sources wrong at their dates; #98's claim on the filters and the makers' account stand with their sources, and the question of what a document keeps of its author stays a question (open block).", "considered": [ { "value": "constructs", "rejected": "the archive does not read Dolma as made by reception; the census takes the paper's account of the corpus as the account" }, { "value": "revises", "rejected": "Dolma is not the archive's own entity; no archive text states it earlier and narrows or withdraws it" } ] }, "translation": [ { "lineage": "what it is: an open English pretraining corpus", "field_claims": [ "F1", "F13", "F17", "F38", "F43", "F54", "F55" ], "received": "F1 (B1, opening): \"Dolma is a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials.\" | F13 (B2, opening): \"Dolma is two things: … Dolma Dataset : an open dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials. … Dolma Toolkit : a high-performance toolkit for curating datasets for language modeling\" | F17 (B3, opening): \"Today, we release our first data artifact in this project — Dolma ¹, a dataset of 3 trillion tokens from a diverse mix of web content, academic publications, code, books, and encyclopedic materials. Openly available for download on the HuggingFace Hub , Dolma is the largest open dataset to date. … For more details, we also release a data sheet (Gebru et al, 2018) as initial documentation.\" | F38 (B4, Abstract): \"To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials.\" | F43 (B5, Dolma: an Open Corpus … [1]): \"In [1], authors from AI2 solve this problem by releasing Data for Open Language Models’ Appetite (Dolma), a fully-open pretraining dataset containing over three trillion tokens of English text\" | F54 (B6, Abstract): \"To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus\" | F55 (B7, Olmo model family): \"In February 2024, 1B and 7B parameter variants of the model were open-sourced, including code, model weights with intermediate snapshots and logs, and contents of their Dolma training dataset (a three-trillion-token English corpus drawn from web text, academic publications, code, books, social media, and encyclopedic sources),\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-01, TLL-02, TLL-04, TLL-10, TLL-11", "archive_claims": [ "TLL-01", "TLL-02", "TLL-04", "TLL-10", "TLL-11" ] }, { "lineage": "versions and size", "field_claims": [ "F2", "F4", "F5", "F12", "F34" ], "received": "F2 (B1, news (2024-04-17)): \"2024-04-17: Dolma v1.7 Release. We have released an updated version of Dolma that we used to train our latest OLMo 7B-v1.7 model.\" | F4 (B1, Versions (table)): \"At the moment, there are six versions of Dolma available: … v1_7 … 2024-04-15 … Used to train OLMo-7B-v1.7 . New sources, more quality filtering, fuzzy deduplication. … v1_6 … 2024-01-31 … v1_6-sample … A smaller sample of Dolma, with roughly 10 billion tokens. … v1_5 … 2023-10-31 … Used to train OLMo-1B . Roughly 3 trillion tokens. … v1 … 2023-08-18 … The first version of Dolma.\" | F5 (B1, Summary Statistics (v1.7), Total): \"Total … 2532.0 … (A subset of total data was used for training of OLMo 7B-v1.7. The token counts are based on the full dataset, whereas taking into account sampling proportion gives the final actual token counts used for training --- 1.715 trillion tokens.)\" | F12 (B1, Summary Statistics (v1.6), Total): \"Total … 11,519 … 4,367 … 2,318 … 3,059\" | F34 (B3, overview figure (image alt text, in B3.html)): \"Overview of Dolma by subset type. Sources include CommonCrawl, C4, peS2o, The Stack, Project Gutenberg, and Wikipedia with a total of 5,245 millions of documents at 5,334 GB, and 3,084 billions of tokens.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-09, TLL-17, TLL-26, CAP-1-03", "archive_claims": [ "TLL-09", "TLL-17", "TLL-26", "CAP-1-03" ] }, { "lineage": "licence and access", "field_claims": [ "F3", "F10", "F16", "F32" ], "received": "F3 (B1, news (2024-04-15)): \"2024-04-15: License Change. We have updated the license of Dolma to ODC-BY .\" | F10 (B1, Licensing Information): \"We are releasing this dataset under the terms of ODC-BY . By using this dataset, you are also bound any license agreements and terms of use of the original data sources.\" | F16 (B2, Dolma Dataset): \"Dolma is licensed under ODC-BY ; see our blog post for explanation.\" | F32 (B3, 6. Releasing Dolma): \"Dolma is released under Ai2's ImpACT license as a medium-risk artifact. Under this license, researchers must: … Provide their contact information and state their intended use case(s) for accessing Dolma; … Disclose the creation of any derivative based on Dolma; … Distribute derivatives under the same restrictions as the ImpACT license; … Agree not to leverage Dolma in a range of prohibited uses, such as military surveillance or generating disinformation.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-24, TLL-25", "archive_claims": [ "TLL-24", "TLL-25" ] }, { "lineage": "what it holds: sources", "field_claims": [ "F6", "F7", "F24", "F29", "F30" ], "received": "F6 (B1, Summary Statistics (v1.7), Source and Provenance columns): \"Refined Web … StarCoder … Reddit … PushShift API … arXiv … RedPajama v1 … StackExchange … Flan … CC News … OpenWebMath … Algebraic Stack … MegaWika … Wikipedia & Wikibooks\" | F7 (B1, Summary Statistics (v1.6), Project Gutenberg row): \"Project Gutenberg … books … 20.4 … 0.056 … 4.0 … 6.0\" | F24 (B3, 3., Web data): \"Our dataset is derived from 24 Common Crawl snapshots collected between 2020–05 to 2023–06. We use the CCNet pipeline to obtain the main content of each web page in plain text form. Further, we also use the C4 dataset, which is obtained from a Common Crawl snapshot collected in April 2019.\" | F29 (B3, 3., Diverse sources): \"For Dolma, we leverage Semantic Scholar's corpus by including papers from peS2o , a subset of 38M permissibly licensed scientific manuscripts. We also include Wikipedia and Project Gutenberg. … For Wikipedia, we use the English and Simplified English subset. For the books in Project Gutenberg, we filter for books that are primarily in English.\" | F30 (B3, 3., Just enough code): \"We derive the code subset of Dolma from The Stack , a collection of permissively-licensed GitHub repositories.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-18, CAP-1-01, O753-01", "archive_claims": [ "TLL-18", "CAP-1-01", "O753-01" ] }, { "lineage": "what of a work's author and source it keeps", "field_claims": [ "F8", "F20" ], "received": "F8 (B1, Summary Statistics (v1.7), header and first row): \"Source … Provenance … New? … Documents (millions) … OLMo tokens (billions) … Sample Proportion … Cutoff Date … Processing … Dolma's CC … Common Crawl via Dolma v1.6\" | F20 (B3, 1. Project Goals, Openness): \"Beyond transparency, open data is critical for research that has become increasingly important with the proliferation of generative models, such as attribution of model output back to pretraining data.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-12, TLL-21, TLL-22, TLL-23, TLL-29, TLL-31, CAP-1-05, D98-02, O186-01, O127-01, O127-02, O127-03, O127-04, O1070-01, O716-01", "archive_claims": [ "TLL-12", "TLL-21", "TLL-22", "TLL-23", "TLL-29", "TLL-31", "CAP-1-05", "D98-02", "O186-01", "O127-01", "O127-02", "O127-03", "O127-04", "O1070-01", "O716-01" ] }, { "lineage": "personal data and removal", "field_claims": [ "F9", "F21", "F33", "F48" ], "received": "F9 (B1, opening): \"Want to request removal of personal data? Use this form to notify us of documents containing PII about a specific user.\" | F21 (B3, 1. Project Goals, Risk Mitigation): \"Risk Mitigation. We should minimize risk posed to individuals while meeting reproducibility and representativeness requirements. For example, we are concerned to what extent content in web crawled data can be traced back to a real-world individual.\" | F33 (B3, 6. Releasing Dolma): \"Finally, as part of our risk mitigation strategy, we created a mechanism to allow the removal of personal data upon request.\" | F48 (B5, Dolma design principles): \"To minimize risk of harm to individuals, one can easily request— using this form —for their data to be removed from the Dolma corpus.\"", "receiving_text": "—", "translation": "untranslated: composed as received", "archive_claims": [] }, { "lineage": "models trained on it, and the training loop", "field_claims": [ "F11", "F50", "F56" ], "received": "F11 (B1, models trained on the dataset): \"Models trained or fine-tuned on allenai/dolma … Browse 241 models trained on this dataset\" | F50 (B5, OLMo): \"OLMo is a set of truly open LLMs that are pretrained on Dolma.\" | F56 (B7, Olmo model family): \"making it the most open state-of-the-art model available.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-14, CAP-1-04, D1695-01, O186-02, O186-03, O1543-01, O1543-02", "archive_claims": [ "TLL-14", "CAP-1-04", "D1695-01", "O186-02", "O186-03", "O1543-01", "O1543-02" ] }, { "lineage": "maker: Ai2 and OLMo", "field_claims": [ "F14", "F18", "F36" ], "received": "F14 (B2, Dolma Dataset): \"It was created as a training corpus for OLMo , a language model from the Allen Institute for AI (AI2).\" | F18 (B3, opening): \"Since March, we at the Allen Institute for AI have been creating OLMo , an open language model to promote the study of large-scale NLP systems.\" | F36 (B3, Notes): \"Fun fact: Dolma stands for \" Data to feed OLMo's Appetite \"\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-03, CAP-1-02", "archive_claims": [ "TLL-03", "CAP-1-02" ] }, { "lineage": "curation: deduplication, masking, decontamination", "field_claims": [ "F15", "F27", "F28", "F31", "F40", "F46", "F47" ], "received": "F15 (B2, Dolma Toolkit): \"High Performance ⚡: Can process billions of documents concurrently thanks to built-in parallelism. … Built-In Taggers 🏷: Includes ready-to-use taggers commonly used to curate datasets such as Gopher , C4 , and OpenWebText . … Fast Deduplication 🗑: Speedy document deduplication using a Rust Bloom filter.\" | F27 (B3, 3., Deduplication): \"First, in common crawl data, we deduplicate pages based on their URL, keeping only one copy of each. Then, we remove duplicate paragraphs within single documents. Both stages use a Bloom filter data structure.\" | F28 (B3, 3., Risk mitigation): \"We detect and mask email addresses, phone numbers, and IP addresses. We also remove content that is detected to be harmful or obscene by fasttext classifiers trained on the Jigsaw dataset .\" | F31 (B3, 3., Decontamination): \"Our decontamination step removes less than 0.001% of the training data by characters, and impacts fewer than 0.02% of documents.\" | F40 (B4, Abstract): \"We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices.\" | F46 (B5, Content filtering): \"To detect PII, a series of regular expression are adopted to find spans of text corresponding to email addresses, IP addresses, and phone numbers. From here, these spans of text are either masked, or the entire document is removed from the corpus (if more than 5 pieces of PII are detected in a single document).\" | F47 (B5, Deduplication): \"In [1], authors perform three stages of deduplication for web text: … URL-based : eliminates web pages that are scraped multiple times. … Document-based : eliminates pages that contain exactly the same text. … Paragraph-based : eliminates individual paragraphs with the same text.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-15", "archive_claims": [ "TLL-15" ] }, { "lineage": "openness, documentation and the toolkit", "field_claims": [ "F19", "F35", "F37", "F39", "F41", "F44" ], "received": "F19 (B3, 1. Project Goals, Openness): \"Openness. Lack of access to pretraining corpora alongside corresponding language models has been a major obstacle for the broader research community. We want to create a dataset that gives other researchers the opportunity to independently create better versions of this dataset, study the relationship between the data and any model trained on it, report any issues they observe when inspecting our data, and critique our curation practices\" | F35 (B3, 5. How Does Dolma Compare With Other Open Datasets?): \"Dolma differentiates itself from other datasets on two key aspects. First, it is significantly larger than other open datasets. Second, it is released under Ai2's impact license, which was designed to balance ease of access with mitigation of potential risk in distributing large datasets.\" | F37 (B4, Abstract): \"Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them.\" | F39 (B4, Abstract): \"We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents.\" | F41 (B4, Abstract): \"Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation.\" | F44 (B5, This overview): \"Unlike prior work in open-source LLMs, Dolma and OLMo are completely transparent and open-source. The process of constructing the Dolma pretraining corpus, which is released under the AI2 impact license , is fully-documented in [1].\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-13, TLL-19, TLL-28, O727-01, O623-01, O1075-01", "archive_claims": [ "TLL-13", "TLL-19", "TLL-28", "O727-01", "O623-01", "O1075-01" ] }, { "lineage": "curation: filtering", "field_claims": [ "F22", "F25", "F26", "F45" ], "received": "F22 (B3, 2. Dataset Design Principles): \"Follow existing practices. … Trust the evaluation suite on interventions it can measure; avoid dramatic shifts when it can't. … Favor decisions that help us (Ai2) with our core research directions. … Take a harms-based approach to risk mitigation.\" | F25 (B3, 3., Quality filtering): \"For example, unigram language models or linear classifiers could be used to select content that broadly resembles Wikipedia pages or books⁴. However, these model-based approaches often have biases that are hard to detect. Following Gopher and Falcon , we opt to use simple heuristics and regular expressions to filter paragraphs.\" | F26 (B3, Notes (note 3)): \"The term \"quality filter\", while widely used in literature, does not appropriately describe the outcome of filtering a dataset. … the filters used in Dolma and other language models efforts select text according to criteria that are inherently ideological .\" | F45 (B5, Quality filtering): \"In [1], authors choose to avoid model-based filtering techniques and rely solely upon heuristic filtering methods.\"", "receiving_text": "—", "translation": "untranslated: composed as received beside D98-01", "archive_claims": [ "D98-01" ] }, { "lineage": "English only", "field_claims": [ "F23", "F49" ], "received": "F23 (B3, 3. How Did We Create Dolma?, English only): \"English only. … We use a fairly permissive threshold, keeping documents that have a likelihood over 50% of being in English. … We recognize that our decision reinforces the assumption of English as the \"default\" language\" | F49 (B5, Language filtering): \"Typically, this language filtering component, which is applied to both web and conversation data in Dolma, eliminates a large amount of source data (e.g., 61.7% of web data is removed).\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-16, TLL-30", "archive_claims": [ "TLL-16", "TLL-30" ] }, { "lineage": "the paper: authors, versions, venue, award", "field_claims": [ "F42", "F51", "F52", "F53" ], "received": "F42 (B4, title block; Comments; Submission history): \"by Luca Soldaini and 35 other authors … Accepted at ACL 2024; Dataset: this https URL Code: this https URL … [v1] Wed, 31 Jan 2024 20:29:50 UTC (7,715 KB) … [v2] Thu, 6 Jun 2024 18:46:40 UTC (6,680 KB)\" | F51 (B6, record): \"Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) … August … Bangkok, Thailand … 15725–15788\" | F52 (B6, record, Award): \"Award: Best Resource Paper Award\" | F53 (B6, authors): \"Luca Soldaini , … Rodney Kinney , … Jesse Dodge , … Kyle Lo\"", "receiving_text": "—", "translation": "untranslated: composed as received beside TLL-27", "archive_claims": [ "TLL-27" ] }, { "lineage": "the name on Wikipedia", "field_claims": [ "F57", "F58" ], "received": "F57 (B7, footer): \"This page was last edited on 13 October 2026, at 05:08\" | F58 (B8, page): \"'''[[Dolma]]''' is a family of stuffed vegetable dishes. … '''Dolma''' may also refer to: … ==People== … * ''[[Dolma (film)]]'', a 2012 Tibetan film\"", "receiving_text": "—", "translation": "untranslated: composed as received", "archive_claims": [] } ], "receiving_texts": "equal-terms case: every field claim composed as received, beside the archive claims it stands with; no receiving text" }, "measure": { "rule": "v0.9 §4.10 (unchanged in v0.10): KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence)", "v0.10": { "surviving": 0, "of": 12, "positions": [ null, null, null, null, null, null, null, null, null, null, null, null ] }, "opening_run": { "rule": "§5.10 (equal-terms case): the field-only KO's sentences as the KO's opening run in their own order", "breach": false }, "blocks": { "rule": "§4.1 (equal-terms case): the KO does not open with the field as a block or close with the archive as one; sentences typed by the ledgers of the claims they cite (F field-only, A archive-only, M both)", "sequence": "MMFFMFFFAMFAFMFMFMAMFFAAFMFFAAAAAFAAAAAAA", "field_only_head": 0, "archive_only_tail": 7, "breach": false }, "max_similarity": [ 0.221, 0.358, 0.452, 0.323, 0.213, 0.327, 0.149, 0.112, 0.454, 0.497, 0.525, 0.477 ], "ko_sentence_1_max_similarity_to_KO_B": 0.268, "v0.9": { "surviving": 0, "of": 12, "positions": [ null, null, null, null, null, null, null, null, null, null, null, null ] }, "key_note": "the counts stored under 'v0.10' and again under 'v0.9', which the page builder reads" }, "homonyms": { "names": [ "Dolma (the stuffed dishes; Wikipedia \"Dolma\")", "Dolma (a given name; Wikipedia \"Dolma (disambiguation)\", People)", "Dolma (2012 Tibetan film)" ], "basis": "No card is a homonym and T composes the corpus (per_dim reached 'same'). On Wikipedia 'Dolma' is the family of stuffed dishes, and the disambiguation page lists people of the name and a 2012 Tibetan film (B8); none is a sense of the corpus's name. 'allenai/dolma' and the paper's title are the census's other names for the entity. The distinct entities below are named by the field, the census or the archive beside the entity and are never senses of its name.", "distinct_entities": [ "the Allen Institute for AI (Ai2), the maker (B1–B3, B7), an organization", "OLMo, the models trained on the corpus (B1, B2, B5, B7)", "OLMoTrace, which traces OLMo outputs to their training data, a /non row of its own (datasets/negative-of-the-negative/v2/rows/olmotrace.json)", "Common Crawl, C4, RefinedWeb, The Stack, StarCoder, peS2o (Semantic Scholar), Project Gutenberg, Wikipedia, Reddit (PushShift), RedPajama, Flan, MegaWika: component sources (B1, B3); Common Crawl and Project Gutenberg rows of this census", "The Pile, a corpus B5 compares, a row of this census", "the Dolma Toolkit, named by its makers as the second thing Dolma is (B2), carried within the entity", "Deep (Learning) Focus and Cameron R. Wolfe, the reviewer of card 5", "the Hamilton and Mimno study (#1695), a third party" ] }, "notes": [ "One address, Google AI Overview (expanded), signed out, incognito, 2026-11-01: dolma-dataset-aio-20261101 (OBS-6dd2d86d5018, ADDR-5a024e0d48d0, AXN:0BAE.CAPTURE.🔽🔩⏪🪐🔃💚), six cards, 171 words in the answer body with a closing offer, 14 markers. A NEW address (CAP-1-10); the census record was compiled the same day and no card cites it.", "Fetching: every card saved verbatim with curl and the census agent 'Mozilla/5.0 (compatible; traininglayerliterature-census)' on 2026-11-01 between 15:13 and 15:15 UTC, raw and text with sha256 in field.json; all six HTTP 200. Card URLs were resolved from T's own inline Google redirect links (HTTP 302, one request each). The Hugging Face card, rate-limited for the census at 04:07 UTC (HTTP 429), returned whole. Wikipedia: no article for the corpus (Dolma_(dataset) 404; Dolma_(corpus) 429 twice, the Wikimedia rate limit, recorded and not retried further); the Ai2 article (B7) and the disambiguation page (B8) are the reference sources.", "Pools (v0.10 §0.3, §3.1(a)): B_D = B1–B6 (54 claims), B_R = B7, B8 (4 claims). The census's own pages are cards 1 and 4, carried once in B_D. T against L_F(B_D) uses B1–B6; B7 and B8 enter L_F(B) and L(B ∪ A) only.", "Census fields a fetched source contradicts or shows out of date (reported, not fixed): (1) operator 'the saved sources name no institution': the dataset card (now readable), the toolkit's pages and Ai2's announcement name Ai2 (F14, F18; card 1 under the Ai2 organization); (2) license 'not recorded': ImpACT at the first release (F32), ODC-BY from 15 April 2024 (F3, F16), users also bound by the original sources' terms (F10); (3) founded '2024 (arXiv v1)' and lifecycle 'later versions … unread': first version 18 August 2023, six versions to v1.7 of 15 April 2024 (F4, F17); (4) unit 'not recorded': the card counts documents (millions) and tokens per source (F5, F12); (5) does_not_answer 'which models other than the authors' own were trained on it': the card lists 241 models trained or fine-tuned on the dataset, among them dmis-lab's UltraBERT (F11); (6) stated_limits 'states no limits': the makers state English only as reinforcing English as the default, filters as 'inherently ideological', and the decontamination's extent (F23, F26, F31); (7) record.name stays unanswered at the document level; the card records provenance per source collection (F8); (8) relations []: the neighbours OLMoTrace (rows/olmotrace.json; traces OLMo outputs to training data), Common Crawl, Project Gutenberg and The Pile (rows of this census; component sources or compared corpora, B3, B5).", "Capture findings the field ledger shows wrong or needing qualification (reported, not fixed): (1) findings[2] ('Version 1.7, the document count and the licences are on no saved source'; 'the award is on no saved source') holds for the census's saved sources; on the cards, v1.7 and ODC-BY are on card 1, ImpACT on cards 3 and 5, the award on card 6. The finding does not name a substitution its own quotation shows: the composition's '~5.2 billion documents totaling over 5,300 GB' are the first release's figures (card 3's '5,245M documents, 5,334 GB', F34) set under version 1.7, whose own table gives 2,532 million documents (F5). (2) The composition's 'PII/toxicity masking' composes toxicity as masked; the makers remove harmful or obscene content and mask personal information (F28); the registry does not record it. (3) findings[1] 'three cards carry the same': cards 2 and 3 name Ai2 by name, card 1's snippet 'AllenAI', card 5 'authors from AI2' (F43).", "T's inline link on 'OLMo' resolves to card 5 (Deep (Learning) Focus), not to an OLMo page; its link on 'Allen Institute for Artificial Intelligence (Ai2)' resolves to card 3, Ai2's Dolma announcement (goto-resolved.txt).", "#98's placing of Dolma among perplexity-based or classifier-based filters is contested by the field (the makers' heuristic quality filtering, B3; the reviewer's 'solely upon heuristic filtering', B5); both composed (P_coh), the question in the open block. The makers do use fastText classifiers for toxicity and language.", "The KO's sentence sequence by ledger ends with seven archive-only sentences (the census's question and the composition at the name, ¶10–¶11, and the loop's archive claims); no field-only opening run (head 0), so §4.1's block test records no breach; recorded as data.", "Homonyms: none on the rail; T composes the corpus (reached 'same'). Wikipedia's 'Dolma' is the dish; its disambiguation lists people and a 2012 film (B8).", "Not run: scripts/build_non.py (only --check), scripts/non_axn.py; panel.json not edited; the session seats the panel entry, mints the AXN (0BAE) and builds the pages." ], "composition": { "source": "this row (composed directly under the brief rebuild/negative-of-the-negative/v010/BRIEF-NEW-ROW-LAYER.md)", "composer": "Claude (Opus 5.5), subagent, 2026-11-01", "spec": "EA-NEGONT-02 v0.10 (#1726)", "defining_material_read": [ { "what": "the specification v0.10", "file": "data/texts/AXN-0B79-text.md", "read": "§0–§11 whole (the procedure), Appendix headings" }, { "what": "the briefs", "file": "rebuild/negative-of-the-negative/v010/BRIEF-NEW-ROW-LAYER.md, BRIEF-NEW-ROW-AUTHORSHIP.md; v09/BRIEF-NEW-ROW-AGGREGATOR.md, BRIEF-NEW-ROW-LABOR.md, BRIEF-NEW-ROW-LEFT.md, BRIEF-NEW-ROW-PROVENANCE.md, BRIEF.md", "read": "whole" }, { "what": "the exemplar row", "file": "datasets/negative-of-the-negative/v2/rows/viaf.json (with its ledgers, reading and config)", "read": "structure field by field" }, { "what": "traininglayerliterature.org layer census record, row", "file": "/home/claude/fleet/traininglayerliterature-org/data/layer-census/census.json (rows[slug == 'dolma'], organizing_question, admission_rule, areas, anchor, schema_note); data/layer-census/rows/dolma.py", "read": "whole" }, { "what": "census sources", "file": "/home/claude/fleet/traininglayerliterature-org/data/layer-census/sources/dolma--card.*, dolma--paper.*, manifest.json", "read": "whole" }, { "what": "the capture", "file": "data/EA-WG-CAPTURES-01.json entries[slug == dolma-dataset-aio-20261101]", "read": "whole (transcript read for T and the delta only)" }, { "what": "field texts", "file": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1–B8", "read": "whole (B5 by its sections on Dolma; B7 at its Olmo passage)" }, { "what": "#186 Who Is Writing for the Machines?", "file": "data/texts/AXN-032C-text.md", "read": "whole" }, { "what": "#1543 Generative Uptake", "file": "data/texts/AXN-0642-text.md", "read": "whole" }, { "what": "#98, #127, #1070, #727, #716, #753, #623, #1075, #1695", "file": "data/texts/AXN-0271, -02B2, -043F, -0278, -025F, -02AB, -01DA, -0444, -0705", "read": "at the candidate loci and their sections (#127 §§ abstract, IV–V)" } ], "cuts": [], "notes": "equal case; KO 41 sentences, 58 of 58 field and 64 of 64 archive claims cited; P body 350 words" }, "composition_v09": { "source": "this row (composed directly under the brief rebuild/negative-of-the-negative/v010/BRIEF-NEW-ROW-LAYER.md)", "composer": "Claude (Opus 5.5), subagent, 2026-11-01", "spec": "EA-NEGONT-02 v0.10 (#1726)", "defining_material_read": [ { "what": "the specification v0.10", "file": "data/texts/AXN-0B79-text.md", "read": "§0–§11 whole (the procedure), Appendix headings" }, { "what": "the briefs", "file": "rebuild/negative-of-the-negative/v010/BRIEF-NEW-ROW-LAYER.md, BRIEF-NEW-ROW-AUTHORSHIP.md; v09/BRIEF-NEW-ROW-AGGREGATOR.md, BRIEF-NEW-ROW-LABOR.md, BRIEF-NEW-ROW-LEFT.md, BRIEF-NEW-ROW-PROVENANCE.md, BRIEF.md", "read": "whole" }, { "what": "the exemplar row", "file": "datasets/negative-of-the-negative/v2/rows/viaf.json (with its ledgers, reading and config)", "read": "structure field by field" }, { "what": "traininglayerliterature.org layer census record, row", "file": "/home/claude/fleet/traininglayerliterature-org/data/layer-census/census.json (rows[slug == 'dolma'], organizing_question, admission_rule, areas, anchor, schema_note); data/layer-census/rows/dolma.py", "read": "whole" }, { "what": "census sources", "file": "/home/claude/fleet/traininglayerliterature-org/data/layer-census/sources/dolma--card.*, dolma--paper.*, manifest.json", "read": "whole" }, { "what": "the capture", "file": "data/EA-WG-CAPTURES-01.json entries[slug == dolma-dataset-aio-20261101]", "read": "whole (transcript read for T and the delta only)" }, { "what": "field texts", "file": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1–B8", "read": "whole (B5 by its sections on Dolma; B7 at its Olmo passage)" }, { "what": "#186 Who Is Writing for the Machines?", "file": "data/texts/AXN-032C-text.md", "read": "whole" }, { "what": "#1543 Generative Uptake", "file": "data/texts/AXN-0642-text.md", "read": "whole" }, { "what": "#98, #127, #1070, #727, #716, #753, #623, #1075, #1695", "file": "data/texts/AXN-0271, -02B2, -043F, -0278, -025F, -02AB, -01DA, -0444, -0705", "read": "at the candidate loci and their sections (#127 §§ abstract, IV–V)" } ], "cuts": [], "notes": "equal case; KO 41 sentences, 58 of 58 field and 64 of 64 archive claims cited; P body 350 words" }, "field_by_origin": { "cards": [ "B1", "B2", "B3", "B4", "B5", "B6" ], "own_pages": [ "B1", "B4" ], "own_pages_note": "the census's two own pages, dolma--card (https://huggingface.co/datasets/allenai/dolma) and dolma--paper (https://arxiv.org/abs/2402.00159), are cards 1 and 4: carried once, as cards, in B_D. The census read dolma--card as HTTP 429; this row's fetch returned it whole (HTTP 200).", "body_named": [], "body_named_note": "T's inline links each resolve to a card (goto-resolved.txt)", "wikipedia": [ "B7", "B8" ], "census_third_party": [], "unreachable": [], "title_only": [], "wikipedia_state": "no article: 'Dolma_(dataset)' raw wikitext HTTP 404 (2026-11-01T15:13:31Z); 'Dolma_(corpus)' HTTP 429 twice (15:13:32Z, 15:14:29Z; Wikimedia rate limit, recorded unread, not retried further); 'Dolma' is the article on the dish (HTTP 200, a homonym); 'Dolma (disambiguation)' (B8, HTTP 200) lists the dish, thirteen people and a 2012 Tibetan film, no dataset. The corpus is stated in the article 'Allen Institute for AI' (B7, last edited 13 October 2026), Olmo model family section. Both are B_R; neither was surfaced by T.", "pools": { "B_D": [ "B1", "B2", "B3", "B4", "B5", "B6" ], "B_R": [ "B7", "B8" ], "note": "v0.10 §0.3, §3.1(a): B_D, every card on the rail; every source T's body names is a card (T's six inline Google redirect links each resolve to a card URL: 'Allen Institute for Artificial Intelligence (Ai2)' to card 3, 'AllenAI Dolma on Hugging Face' to card 1, 'OLMo' to card 5, 'Dolma ACL Research Paper' to card 6, the markers to cards 1, 2 and 4), so no body-named source stands apart. The census's two own pages are cards 1 (dolma--card) and 4 (dolma--paper), carried once in B_D. B_R: the Wikipedia reference sources (B7, the Ai2 article at its Dolma passage; B8, the disambiguation page as the record of absence). The census read no third-party source for the row. T against L_F(B_D) uses B1–B6 alone." }, "sources": [ { "id": "B1", "origin": "card", "also_origin": "own_page", "pool": "B_D", "role": "Overview card 1 (and the census's own page dolma--card: one page, carried once, in B_D)", "card": "Hugging Face — allenai/dolma · Datasets at Hugging Face", "url": "https://huggingface.co/datasets/allenai/dolma", "kind": "Overview card (the dataset card on the Hugging Face Hub, under the allenai namespace); the census's own page dolma--card", "files": { "B1.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1.html", "sha256": "ba6f8001031a5e0df08cf736088e1985fdc52edc9bd270f534ce412065c7aedd" }, "B1.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1.txt", "sha256": "bd31e420de09639648bf5dfe543ed804f42f4a9a2fc4e11fafc87f89e1318852" }, "B1-census.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1-census.html", "sha256": "967409f981642b7ce52d05e80b51fc1dd8dfcfa8cbbea9623106b2913ff7e467" }, "B1-census.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1-census.txt", "sha256": "f56acaeffb2ad04bd6efb118c4a4443c271a2d891bd3e7e740beac86676c2930" } }, "fetched": { "at": "2026-11-01T15:13:22Z", "http": 200, "url": "https://huggingface.co/datasets/allenai/dolma", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1.html", "raw_sha256": "ba6f8001031a5e0df08cf736088e1985fdc52edc9bd270f534ce412065c7aedd", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B1.txt" ], "census_copy": { "src": "dolma--card", "http": "429", "fetched": "2026-11-01T04:07:51Z", "sha256": "967409f981642b7ce52d05e80b51fc1dd8dfcfa8cbbea9623106b2913ff7e467", "agrees": "no: the census's fetch at 04:07:51Z was rate-limited (HTTP 429, 'We had to rate limit your IP'); this fetch at 15:13:22Z returned the card whole (HTTP 200)" } }, "names_operator": "yes ('Ai2' as the namespace's organization; 'our latest OLMo 7B-v1.7 model'; the BibTeX authors)", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "URL from T's inline link 'AllenAI Dolma on Hugging Face' and card 1's title (resolved from T's inline Google redirect links, one request each with the census agent at 2026-11-01T15:13:11–13Z (HTTP 302; goto-resolved.txt)). The census's own page for the entity; the census read it as HTTP 429 and recorded it unread. The card carries the versions table, the v1.7 and v1.6 summary statistics, the download instructions, the licence and the models trained on the dataset.", "surfaced_by_T": [ { "slug": "dolma-dataset-aio-20261101", "card": 1 } ], "card_title": "allenai/dolma · Datasets at Hugging Face", "card_snippet": "Overview: AllenAI Dolma is an open corpus dataset containing 3 trillion tokens (version 1.7 utilizing 1.715 trillion actual training tokens) designed for langua..." }, { "id": "B2", "origin": "card", "pool": "B_D", "role": "Overview card 2", "card": "GitHub Pages documentation — dolma | Data and tools for generating and inspecting OLMo pre-training data.", "url": "https://allenai.github.io/dolma/", "kind": "Overview card (the GitHub Pages site of the Dolma repository: the dataset and the Dolma Toolkit)", "files": { "B2.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B2.html", "sha256": "9b7b9191b66c1ae1b7594ce6e3806ecd00384bf261f525a00594603034879822" }, "B2.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B2.txt", "sha256": "1ceed892a982dccdf7e33bfcee7aabe4565edad005fbcaf84577499bd50bb14b" } }, "fetched": { "at": "2026-11-01T15:13:22Z", "http": 200, "url": "https://allenai.github.io/dolma/", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B2.html", "raw_sha256": "9b7b9191b66c1ae1b7594ce6e3806ecd00384bf261f525a00594603034879822", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B2.txt" ] }, "names_operator": "yes ('a language model from the Allen Institute for AI (AI2)')", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "URL resolved from T's second citation link (resolved from T's inline Google redirect links, one request each with the census agent at 2026-11-01T15:13:11–13Z (HTTP 302; goto-resolved.txt)); the page's is the card's title verbatim.", "surfaced_by_T": [ { "slug": "dolma-dataset-aio-20261101", "card": 2 } ], "card_title": "dolma | Data and tools for generating and inspecting OLMo pre-training data.", "card_snippet": "Dataset: Open 3-trillion-token corpus (web content, academic publications, code, books, encyclopedic materials) created as training data for OLMo by AI2; availa..." }, { "id": "B3", "origin": "card", "pool": "B_D", "role": "Overview card 3", "card": "Allen Institute for Artificial Intelligence — Ai2 Dolma: 3 trillion token open corpus for language model pretraining | Ai2", "url": "https://allenai.org/blog/dolma-3-trillion-tokens-open-llm-corpus-9a0ff4b8da64", "kind": "Overview card (Ai2's announcement of the first release, 18 August 2023, by Luca Soldaini)", "files": { "B3.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B3.html", "sha256": "3ca32b1bc011de1c60d337dc2b4e8f9e7c5a59ce1987697024be9c130fefe688" }, "B3.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B3.txt", "sha256": "006aa43afa5bb83414023bf4eea4aec411706764f8e2ec255325992915f08172" } }, "fetched": { "at": "2026-11-01T15:13:23Z", "http": 200, "url": "https://allenai.org/blog/dolma-3-trillion-tokens-open-llm-corpus-9a0ff4b8da64", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B3.html", "raw_sha256": "3ca32b1bc011de1c60d337dc2b4e8f9e7c5a59ce1987697024be9c130fefe688", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B3.txt" ], "note": "the overview figure's alt text (5,245 millions of documents, 5,334 GB, 3,084 billions of tokens) is in B3.html, an attribute stripped from B3.txt; its quote is checked against B3.html" }, "names_operator": "yes ('we at the Allen Institute for AI'; '© The Allen Institute for Artificial Intelligence')", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "URL resolved from T's inline link 'Allen Institute for Artificial Intelligence (Ai2)' (resolved from T's inline Google redirect links, one request each with the census agent at 2026-11-01T15:13:11–13Z (HTTP 302; goto-resolved.txt)); the page's title is card 3's verbatim. The card snippet's '5,245M documents, 5,334 GB' renders the overview figure's alt text.", "surfaced_by_T": [ { "slug": "dolma-dataset-aio-20261101", "card": 3 } ], "card_title": "Ai2 Dolma: 3 trillion token open corpus for language model pretraining | Ai2", "card_snippet": "Dataset: Dolma (3 trillion tokens, 5,245M documents, 5,334 GB) developed by Ai2 for OLMo pretraining. Sources: CommonCrawl, C4, peS2o, The Stack, Project Gutenb..." }, { "id": "B4", "origin": "card", "also_origin": "own_page", "pool": "B_D", "role": "Overview card 4 (and the census's own page dolma--paper: one page, carried once, in B_D)", "card": "arXiv — [2402.00159] Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", "url": "https://arxiv.org/abs/2402.00159", "kind": "Overview card (the paper's arXiv abstract page, v2); the census's own page dolma--paper", "files": { "B4.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B4.html", "sha256": "1203d77da94d4612ee60d200ff18f63ce0f75d4100ef6ae5301d972d4f445946" }, "B4.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B4.txt", "sha256": "81726256243b7fe6a3c735bc23f09effda44b7eb759e2d2cc3c2dd408b84dafc" }, "B4-census.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B4-census.html", "sha256": "1203d77da94d4612ee60d200ff18f63ce0f75d4100ef6ae5301d972d4f445946" }, "B4-census.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B4-census.txt", "sha256": "81726256243b7fe6a3c735bc23f09effda44b7eb759e2d2cc3c2dd408b84dafc" } }, "fetched": { "at": "2026-11-01T15:13:23Z", "http": 200, "url": "https://arxiv.org/abs/2402.00159", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B4.html", "raw_sha256": "1203d77da94d4612ee60d200ff18f63ce0f75d4100ef6ae5301d972d4f445946", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B4.txt" ], "census_copy": { "src": "dolma--paper", "http": "200", "fetched": "2026-11-01T04:07:51Z", "sha256": "1203d77da94d4612ee60d200ff18f63ce0f75d4100ef6ae5301d972d4f445946", "agrees": "yes: byte-identical (same sha256)" } }, "names_operator": "authors only ('by Luca Soldaini and 35 other authors'); no institution", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "URL resolved from T's first citation link (resolved from T's inline Google redirect links, one request each with the census agent at 2026-11-01T15:13:11–13Z (HTTP 302; goto-resolved.txt)). The census's one read source for the row; the census's saved copy and this fetch are byte-identical.", "surfaced_by_T": [ { "slug": "dolma-dataset-aio-20261101", "card": 4 } ], "card_title": "[2402.00159] Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", "card_snippet": "Overview: Dolma is an open-source, three-trillion-token English corpus designed for language model pretraining research, addressing the scarcity of publicly ava..." }, { "id": "B5", "origin": "card", "pool": "B_D", "role": "Overview card 5", "card": "Deep (Learning) Focus — Dolma, OLMo, and the Future of Open-Source LLMs", "url": "https://cameronrwolfe.substack.com/p/dolma-olmo-and-the-future-of-open", "kind": "Overview card (Cameron R. Wolfe's newsletter overview of the Dolma and OLMo papers, posted 2024-02-20)", "files": { "B5.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B5.html", "sha256": "e46900187043f42865f10f47653e6d23602463a8e31f0101e6cef4e31f55e7d5" }, "B5.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B5.txt", "sha256": "ad92d43ad246e79f9e86e65ee3f05c0c849770534913ecefe3bbe0481d6e89c7" } }, "fetched": { "at": "2026-11-01T15:13:24Z", "http": 200, "url": "https://cameronrwolfe.substack.com/p/dolma-olmo-and-the-future-of-open", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B5.html", "raw_sha256": "e46900187043f42865f10f47653e6d23602463a8e31f0101e6cef4e31f55e7d5", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B5.txt" ], "post_date": "2024-02-20T10:33:01.466Z (the page's post_date)" }, "names_operator": "yes ('authors from AI2')", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "URL resolved from T's citation link and from the inline link on 'OLMo' (the same redirect; resolved from T's inline Google redirect links, one request each with the census agent at 2026-11-01T15:13:11–13Z (HTTP 302; goto-resolved.txt)). A third-party account of the paper, attributed claims ('In [1], authors ...').", "surfaced_by_T": [ { "slug": "dolma-dataset-aio-20261101", "card": 5 } ], "card_title": "Dolma, OLMo, and the Future of Open-Source LLMs", "card_snippet": "Overview: Dolma (3T-token open pretraining corpus) and OLMo (fully open-source LLM suite) demystify language model development by openly releasing training data..." }, { "id": "B6", "origin": "card", "pool": "B_D", "role": "Overview card 6", "card": "ACL Anthology — Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - ACL Anthology", "url": "https://aclanthology.org/2024.acl-long.840/", "kind": "Overview card (the paper's ACL Anthology record: authors, abstract, proceedings, pages, award)", "files": { "B6.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B6.html", "sha256": "08bbcf3ad027b9f6b3c0cb92827ac6dab74ab3f2f33758f567c658a60d3cabfc" }, "B6.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B6.txt", "sha256": "781aa6ce3c0caa4af1bdf5e4a10cac8d64e487297104f44e9fb119773ef7f16f" } }, "fetched": { "at": "2026-11-01T15:13:24Z", "http": 200, "url": "https://aclanthology.org/2024.acl-long.840/", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B6.html", "raw_sha256": "08bbcf3ad027b9f6b3c0cb92827ac6dab74ab3f2f33758f567c658a60d3cabfc", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B6.txt" ] }, "names_operator": "authors only (36 names); no institution", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "URL resolved from T's inline link 'Dolma ACL Research Paper' (resolved from T's inline Google redirect links, one request each with the census agent at 2026-11-01T15:13:11–13Z (HTTP 302; goto-resolved.txt)). Its abstract is the arXiv abstract word for word (one lineage, B4 the earlier, 31 January 2024).", "surfaced_by_T": [ { "slug": "dolma-dataset-aio-20261101", "card": 6 } ], "card_title": "Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - ACL Anthology", "card_snippet": "Resource: Dolma (ACL 2024 Best Resource Paper Award). Corpus: Open English dataset containing three trillion tokens for language model pretraining research. Con..." }, { "id": "B7", "origin": "wikipedia", "pool": "B_R", "role": "Wikipedia: Dolma has no article; the corpus is stated in the article on its maker, 'Allen Institute for AI', at the passage on the OLMo family. Not surfaced by T", "card": "Wikipedia — Allen Institute for AI (the passage on Dolma)", "url": "https://en.wikipedia.org/wiki/Allen_Institute_for_AI", "kind": "Wikipedia article on Ai2, read at its Olmo model family section", "files": { "B7.html": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B7.html", "sha256": "b7669180cb2106832182baf31da7dd62eee1554dfb3345db164af081782e61f9" }, "B7.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B7.txt", "sha256": "3083e6a9400a1f4bd09452e257e147f9b907341d87bad76f6b00f15a0eeb660c" }, "B7.wiki": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B7.wiki", "sha256": "87b14f4b9a1f79c897d5662b7b56ce98ca992672a747a682624b2a3d1cb57321" } }, "fetched": { "at": "2026-11-01T15:15:19Z", "http": 200, "url": "https://en.wikipedia.org/wiki/Allen_Institute_for_AI", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B7.html", "raw_sha256": "b7669180cb2106832182baf31da7dd62eee1554dfb3345db164af081782e61f9", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B7.txt" ], "wikitext": { "url": "https://en.wikipedia.org/w/index.php?title=Allen_Institute_for_AI&action=raw", "http": 200, "at": "2026-11-01T15:14:35Z", "sha256": "87b14f4b9a1f79c897d5662b7b56ce98ca992672a747a682624b2a3d1cb57321" } }, "names_operator": "yes (the article is on the operator)", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "No article for the corpus: 'Dolma_(dataset)' raw HTTP 404 (15:13:31Z); 'Dolma_(corpus)' HTTP 429 twice (15:13:32Z, 15:14:29Z, Wikimedia rate limit; not retried further); 'Dolma' is the article on the dish (HTTP 200, W-Dolma.wiki); 'Dolma (disambiguation)' lists no dataset (B8). The Ai2 article is the encyclopedia's one statement of the corpus found; last edited 13 October 2026. A reference source (§0.3): it shows nothing of what T had before it.", "surfaced_by_T": [] }, { "id": "B8", "origin": "wikipedia", "pool": "B_R", "role": "Wikipedia: 'Dolma (disambiguation)', the record of the name's senses on Wikipedia, none of them the corpus. Not surfaced by T", "card": "Wikipedia — Dolma (disambiguation) (the record of absence)", "url": "https://en.wikipedia.org/w/index.php?title=Dolma_(disambiguation)&action=raw", "kind": "Wikipedia disambiguation page (raw wikitext)", "files": { "B8.wiki": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B8.wiki", "sha256": "68d67621337c934aa530276c1c214f721a2f7003d5d3cd82d0dc820bbc0cbd75" }, "B8.txt": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B8.txt", "sha256": "68d67621337c934aa530276c1c214f721a2f7003d5d3cd82d0dc820bbc0cbd75" }, "W-Dolma.wiki": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/W-Dolma.wiki", "sha256": "1728beffb5b484ab4d5f1f1e86bb57328071c5b14a51c3c640904c2bd3b88269" }, "W-Dolma_(dataset).wiki": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/W-Dolma_(dataset).wiki", "sha256": "95c3fc2cc9fb2a6f55704845850a70217992db01b2b987901169e38d6abc60cd" }, "W-Dolma_(corpus).wiki": { "path": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/W-Dolma_(corpus).wiki", "sha256": "33b02e3b35f9b84ad543533dba9932c83e1e2c76fa5798339e0d47393a77a00a" } }, "fetched": { "at": "2026-11-01T15:14:03Z", "http": 200, "url": "https://en.wikipedia.org/w/index.php?title=Dolma_(disambiguation)&action=raw", "agent": "Mozilla/5.0 (compatible; traininglayerliterature-census)", "raw": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B8.wiki", "raw_sha256": "68d67621337c934aa530276c1c214f721a2f7003d5d3cd82d0dc820bbc0cbd75", "text": [ "datasets/negative-of-the-negative/v2/intake/dolma-20261101/B8.txt" ], "absence": [ { "title": "Dolma_(dataset)", "http": 404, "at": "2026-11-01T15:13:31Z", "file": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/W-Dolma_(dataset).wiki" }, { "title": "Dolma_(corpus)", "http": 429, "at": [ "2026-11-01T15:13:32Z", "2026-11-01T15:14:29Z" ], "file": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/W-Dolma_(corpus).wiki", "note": "Wikimedia rate limit, twice with the same agent; recorded unread, not retried further" }, { "title": "Dolma", "http": 200, "at": "2026-11-01T15:13:32Z", "file": "datasets/negative-of-the-negative/v2/intake/dolma-20261101/W-Dolma.wiki", "note": "the article on the dish ('Stuffed dishes in Middle Eastern cuisine'), a homonym" } ] }, "names_operator": "no", "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, traininglayerliterature, Johannes Sigil, EA-WG-CAPTURES-01 in the saved text, card title or snippet)", "note": "B8.txt is the wikitext as text (the page is wikitext only). The disambiguation lists the dish, thirteen people, a 2012 Tibetan film; no dataset or corpus.", "surfaced_by_T": [] } ] }, "derivatives": { "kernel": { "status": "composer-written", "derived_from": "the current composition, written by its composer under the brief", "note": "§8.2 derives the kernel mechanically from the frozen plan; this one was written by the composer, mechanical derivation pending." }, "delta": { "status": "current", "derived_from": "the current composition" } } }