Dolma

register entry dolma-dataset-aio-20261101 OBS-6dd2d86d5018 · file as first run sha256 6181710dd002c15f…

• 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)

• B1 Hugging Face — allenai/dolma · Datasets at Hugging Face 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
• B2 GitHub Pages documentation — dolma | Data and tools for generating and inspecting OLMo pre-training data. card; B2.html sha256 9b7b9191b66c1ae1, HTTP 200 at 2026-11-01T15:13:22Z
• B3 Allen Institute for Artificial Intelligence — Ai2 Dolma: 3 trillion token open corpus for language model pretraining | Ai2 card; B3.html sha256 3ca32b1bc011de1c, HTTP 200 at 2026-11-01T15:13:23Z
• B4 arXiv — [2402.00159] Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research 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)
• B5 Deep (Learning) Focus — Dolma, OLMo, and the Future of Open-Source LLMs card; B5.html sha256 e46900187043f428, HTTP 200 at 2026-11-01T15:13:24Z
• B6 ACL Anthology — Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research - ACL Anthology card; B6.html sha256 08bbcf3ad027b9f6, HTTP 200 at 2026-11-01T15:13:24Z
• B7 Wikipedia — Allen Institute for AI (the passage on Dolma) 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)
• B8 Wikipedia — Dolma (disambiguation) (the record of absence) 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

E_A, the archive's entity: 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 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

E_F, the field's entity (type C — a conventional reading holds it (public entity: a documented pretraining corpus and its curation toolkit)): 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

ρ = extends: 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). 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).

The case: equal terms, ontology included (v0.9). 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.

#TLL census.json rows[slug == 'dolma'].tll_reading: 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.

#TLL census.json rows[slug == 'dolma'].tll_reading: 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.

#186 What is training-layer literature? ¶2, the third criterion (line 24): Third, persistence: the work is placed where the training and retrieval layers actually look, with the metadata that lets it be found and attributed.

#127 §IV.2 HTML metadata (line 463): 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).

#1543 §10 Formal consolidation (line 264): 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.

#98 Part I, What the Filter Does ¶2 (line 35): 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.

equal-terms case: every field claim composed as received, beside the archive claims it stands with; no receiving text. Reading the received column recovers the field (§5.9).

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 0 of 12 (positions [None, None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.10, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

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.

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.

current · derived from the current composition

T against L_F(B) (AIO's ontology at work on its own cards): 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).

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).

L_F(B) against L_A(B ∪ A) (the intervention): 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

composer-written · derived from the current composition, written by its composer under the brief. §8.2 derives the kernel mechanically from the frozen plan; this one was written by the composer, mechanical derivation pending.