OLMoTrace

register entry olmotrace-aio-20261029 OBS-3fd20ae244c0 · file as first run sha256 cd5b1c0c59328025…

• T1-1 definition: 'an open-source system developed by the Allen Institute for Artificial Intelligence (Ai2) that traces language model outputs back to their multi-trillion-token training datasets in real time' (two markers)
• T1-2 real-time tracing: 'After a model generates a response in the Ai2 Playground, users can click the Show OLMoTrace button to highlight verbatim spans of text.' (two markers)
• T1-3 source retrieval: 'The tool matches output segments against training corpora (such as the Dolma datasets) and ranks up to 10 context documents using BM25 and unigram probabilities.' (two markers)
• T1-4 underlying tech: 'It uses an extended version of infini-gram with a parallel algorithm that reduces lookup complexity to \(O(L \log N)\).' (one marker)
• T1-5 supported models: 'OLMo 2 32B Instruct', 'OLMo 2 13B Instruct', 'OLMoE 1B 7B Instruct', 'OLMo 3 family (including OLMo 3-Think 32B)' (two markers)
• T1-6 fact-checking: 'Fact-Checking: Verify whether specific claims in a model's answer are backed by real training documents or generated from general parameters.' (two markers)
• T1-7 hallucination diagnosis: 'Hallucination Diagnosis: Spot where a model mixes up facts or memorized text from source files.' (two markers)
• T1-8 auditing: 'Auditing & Transparency: Inspect the data recipes and pretraining paths for regulatory compliance or research.' (two markers)
• T1-9 closer: 'Would you like to know how to access the open-source code for OLMoTrace on GitHub/Hugging Face, or are you looking for details on a specific OLMo model version?'

• B1 Allen Institute for Artificial Intelligence — Going beyond open data – increasing transparency and trust in language models with OLMoTrace | Ai2 card; B1.html sha256 39e59dda3ba78e0b, HTTP 200 at 2026-10-29T17:07:13Z
• B2 arXiv — OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training ... card; B2.html sha256 cd982680a1ca4c02, HTTP 200 at 2026-10-29T17:07:17Z; byte-identical to the census copy olmotrace--abs
• B3 ACL Anthology — Tracing Language Model Outputs Back to Trillions of Training Tokens card; B3.html sha256 71292b784f0dc598, HTTP 200 at 2026-10-29T17:07:19Z; PDF sha256 fcac45353235b662
• B4 X — Olmo 3 is wired into OlmoTrace in the Ai2 Playground, so you don't just ... - X unreachable: https://x.com/allen_ai HTTP 200; https://x.com/search?q=%22wired%20into%20OlmoTrace%22 HTTP 200; card title and snippet only
• B5 Techzine Global — Ai2 introduces OLMoTrace: a peek inside LLM's - Techzine Global unreachable: https://www.techzine.eu/news/analytics/130296/ai2-introduces-olmotrace-a-peek-inside-llms/ HTTP 403; https://www.techzine.eu/news/analytics/130296/ai2-introduces-olmotrace-a-peek-inside-llms/ HTTP 403; https://www.techzine.eu/news/analytics/130296/ai2-introduces-olmotrace-a-peek-inside-llms/ HTTP 403; card title and snippet only
• B6 The AI Economy | Ken Yeung — Meet OLMoTrace: The Tool That Shows Where AI Gets Its Info card; B6.html sha256 bf021478cec54cb8, HTTP 200 at 2026-10-29T17:07:23Z
• B7 arXiv — OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens (HTML, v2) own_page; census copy olmotrace--paper (read 2026-10-29, HTTP 200); B7.html sha256 b1b16ba583eeb0db
• B8 Wikipedia — OLMoTrace (no article) no article: raw HTTP 404, rendered HTTP 404 (after one 429), search no results; sha256 92d530a8e9a8411f

E_A, the archive's entity: the same tool, in the provenance census record and the archive's deposits: a tracing tool in the lineage of training data that records provenance, its object a span of model output and the training documents that contain it verbatim, its relation a verbatim match, its carrier the full training corpus, with the authors' limit as its negative tags (not causes; not a citation system); a member of the class of training-data attribution, which operates on the training/output relation where PER operates on the output/reader relation (#725), traces partially and at cost without recovering authorship (#1714), stands as one post-hoc detection layer in isolation (#127), holds similarity as exploratory evidence (#1075), and supplies the training-data provenance records parametric claims need (#95); whose composition at its own name converts the matching relation into a citation relation (Capture Registry, 2026-10-29) PE-01 PE-02 PE-09 PE-03 PE-04 PE-05 PE-15 PE-16 O725-01 O1714-01 O127-01 O1075-02 O95-01 CAP-1-09; defining papers #725, #95, #1075, #127, #1714

E_F, the field's entity (type C — a conventional reading holds it (public entity: a tracing tool and the paper that presents it)): the system of the Allen Institute for AI with the University of Washington, UC Berkeley and Stanford (Liu, Blanton and 29 others; arXiv:2504.07096, 9 April 2025, v2 8 July 2025; ACL 2025 System Demonstrations, Best Demo) that finds and shows, in the Ai2 Playground and in real time, verbatim matches between long, unique spans of a language model's output and documents in the model's full training data (for OLMo-2-32B-Instruct 3.2 billion documents, 4.6 trillion tokens), over an extended infini-gram index, retrieving up to 10 snippets per span ranked by BM25; its purpose to explore where a model may have learned a word sequence, its uses shown on fact checking (a possibility, conditional on the user's verification of the sources), creativity and math; its stated limit that the documents are not causes of the output, nor supporting evidence or citations for it; set apart from retrieval-augmented generation, Google AI Overview among RAG's examples; applicable to any model whose full training data its provider has access to; open source

ρ = extends: The archive reads OLMoTrace as its authors present it and adds, on the field's own object: the relation the class operates on (training/output) against the one PER measures (output/reader); the cost and partiality of tracing and what it does not recover; its isolation as one survival layer; the evidentiary grade of similarity; the parametric claims it is companion infrastructure for; the census's answer, field by field, to what provenance it carries; and the composition layer's conversion, at its name, of a match into support. O725-01 O1714-01 O1714-02 O127-01 O1075-01 O1075-02 O95-01 PE-21 PE-03 PE-04 CAP-1-04 CAP-1-09. Guard: ρ = extends claims an addition and no more (§1.2a). It does not make the field's sources wrong at their dates. Where an archive claim stands against a field claim (PE-17 against F17/F19, open or accessible training data), both are composed under P_coh within the one ontology. The archive's reading of fact checking at the composition layer (CAP-1-04, CAP-1-09) is composed beside the field's own statements of the use (F21–F29) and of the limit (F34–F37), not in their place.

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The provenance census record describes OLMoTrace from its authors' own pages, each field with its basis ('operator'), and answers for it the census's question of what provenance a system carries, through which operation, on whose assertion, with what evidence of survival; the capture record checks the composition against those pages (CAP-1-01–CAP-1-02). No deposit names the tool. The deposits bear on its class, tracing model outputs to training data, and add senses and limits on the field's own object: the training/output relation against PER's output/reader relation (#725), partial tracing at cost that recovers no authorship (#1714), the post-hoc detection layer developed in isolation (#127), similarity as exploratory evidence only (#1075), and the parametric claims such records are companion infrastructure for (#95). #725 states its own relation to the frameworks it places in these terms ('Extension and completion, not substitution or critique'). No defining paper contests the field's ontology of the tool or supplies one the field lacks: ρ = extends, the equal-terms case (§0.9, §1.2a). The archive's reading of the composition, a matching relation converted into a citation relation (CAP-1-09), takes the authors' limit as given and is the field's own distinction carried to the composition layer.

#725 packet header, Relation: **Relation:** Extension and completion, not substitution or critique

#725 §4 table, row 'Data attribution (TRAK, influence functions)': | Data attribution (TRAK, influence functions) | Which training examples shaped a specific output | Adjacent eval method | Operates on training/output relation; PER operates on output/reader relation |

#1714 §9.2 s4–5: The attribution literature shows that partial tracing is possible at cost: influence-function methods estimate which training examples most affect a given output (Grosse et al. 2023). Neither recovers authorship, and both show that what survives depends on what is recorded and computed.

#95 the typology of source-dependent claims: Parametric claims are drawn from the model's training-time parameters rather than runtime retrieval, often without surface signal indicating which sources materially contributed; PER applies to these only indirectly, and addressing them at the substrate level requires companion infrastructure (training-data provenance records, watermarking, content credentials).

#PE census.json organizing_question: What provenance does this system carry, through which operation, on whose assertion, with what evidence of survival?

#PE census.json rows[slug == 'olmotrace'].provenance_object: a span of a language model's output and the training documents that contain it verbatim

#capture-registry olmotrace-aio-20261029, analysis s2: The composition turns a trace of verbatim overlap into evidence for a claim, the conversion of a matching relation into a citation relation that the paper names and refuses.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: 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): v0.9 0 of 14 (positions [None, None, None, None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

OLMoTrace is, in its authors' words, "the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time" (arXiv:2504.07096), introduced by Ai2 on 9 April 2025 as a feature of the Ai2 Playground (Ai2 blog). It finds and shows verbatim matches between segments of a model's output and documents in the training corpora, highlighting long and unique spans that occur verbatim at least once in the training data (arXiv abstract; Ai2 blog), and it retrieves by lexical overlap on the heuristic that such documents are likely to be influential (the paper, §6). It matches against the whole of a model's training data, for OLMo-2-32B-Instruct 3.2 billion documents and 4.6 trillion tokens (the paper; Ai2 blog). Spans are found over an infini-gram index at O(L * log N) (Ai2 blog), up to 10 enclosing snippets are retrieved per kept span and the documents ranked by BM25, a score that conveys "some notion of relevance" (the paper; Ai2 blog), and the user starts it with the "Show OLMoTrace" button below an answer (Techzine, snippet only). The paper measures 4.46 seconds per query and a first displayed document of 1.90 on a 0–3 relevance scale, roughly "on the same topic" (the paper). At launch it served OLMo 2 32B and 13B Instruct and OLMoE 1B 7B Instruct and can be applied to any model whose full training data the provider has access to (Ai2 blog; the paper); Ai2's post on X gives the OLMo 3 family (X, snippet only), and Liu expects closed-model providers to hesitate (Ken Yeung). Its stated purpose is to explore where models "may have learned to generate certain word sequences", verbatim matching taken as the most direct connection (the paper; Ai2 blog), at an early stage of exploring the tool (Ai2 blog). Its authors show it on fact checking, hallucination and creativity, with case studies on fact checking, "creative" expressions and math (the paper), a memorized math expression and a false knowledge cutoff traced to post-training data (Ai2 blog), and Ai2's leaders tie it to hallucination scores and to auditability (Ken Yeung). Fact checking is stated as a possibility, users "may be able to fact-check the statement against its training data", where context and source help a more informed judgment because some sources are unreliable (the paper, both versions), and confidence follows only if users personally verify that the documents come from trusted sources (Ai2 blog); the newsletter calls it fact-checking, and Liu's launch statement says users can ensure the data supporting their outputs is "trustworthy and verifiable" (Ken Yeung). The authors' limit: "The retrieved documents should not be interpreted as having a causal effect on the LM output, or as supporting evidence or citations for the LM output" (the paper, arXiv v2 and ACL 2025); Ai2 told the newsletter the documents are associations, not the definitive source and not indicators of authority (Ken Yeung). The paper sets it apart from retrieval-augmented generation, Google AI Overview among RAG's examples: it retrieves post hoc and does not intervene in generation (the paper; Liu in Ken Yeung), and Gemini's double-check is less useful because Search is not Gemini's training data (the paper). Its 31 authors are at the Allen Institute for AI, the University of Washington, UC Berkeley and Stanford (arXiv; the paper); it is "publicly available and fully open-source", its core under Apache 2.0, the openness Ai2 calls its commitment and Farhadi its condition (arXiv; the paper; Ai2 blog; Ken Yeung). Submitted 9 April 2025 and revised 8 July, it appeared in ACL 2025's System Demonstrations with the Best Demo award (arXiv; ACL Anthology); its authors name the easier exposure of problematic training content as a risk (the paper), and the newsletter finds it gives no analytics for improving data, "generating awareness" (Ken Yeung). Wikipedia has no article on it (Wikipedia, 2026-10-29).

OLMoTrace is a tracing tool in the census's lineage of training data, attaching provenance to "a span of a language model's output and the training documents that contain it verbatim" (provenance census record), and its authors present it as "the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time" (arXiv:2504.07096). Ai2 introduced it on 9 April 2025 as a feature of the Ai2 Playground (Ai2 blog), and the census answers its organizing question for it, "What provenance does this system carry, through which operation, on whose assertion, with what evidence of survival?", field by field (census record). The relation it carries is a verbatim match: it "finds and shows verbatim matches between segments of language model output and documents in the training text corpora" (arXiv abstract), highlighting long and unique spans found at least once in the training data (Ai2 blog), which the census records as its one relation type, its operation (model generation traced back to training text) and its unit, "a maximal matching span of model output" (census record). It retrieves "by lexical overlap, with the heuristic that such training examples are likely to be influential for the given output" (the paper, §6); for the class of training-data attribution, which treats model behaviour as counterfactually dependent on the examples that entered training, the archive's protocol ranks exact paired retraining first, validated approximation second, and "unvalidated similarity or intuition only as exploratory evidence" (#1075). Its carrier is the model's full training corpus over an extended infini-gram index (census record): for OLMo-2-32B-Instruct, 3.2 billion documents and 4.6 trillion tokens across pre-, mid- and post-training, all of them matched (the paper; Ai2 blog; census record). Span finding runs at O(L * log N) (Ai2 blog); up to 10 enclosing snippets are retrieved per kept span and ranked by BM25 (the paper), a score that conveys "some notion of relevance" (Ai2 blog), behind a "Show OLMoTrace" button below the model's answer (Techzine, snippet only). Per query the tracing takes 4.46 seconds on average, and the first document it displays averages 1.90 on a 0–3 relevance scale, roughly "on the same topic" as the output, in the paper's own measurement (the paper). At launch it served OLMo 2 32B and 13B Instruct and OLMoE 1B 7B Instruct and "can be applied to any language model for which you have access to the training data" (Ai2 blog), "as long as the service provider has access to its full training data" (the paper); Ai2's post on X gives the OLMo 3 family (X, snippet only), which the Capture Registry finds the census record does not yet carry and is to carry from Ai2's own page (Capture Registry, 2026-10-29). The census tags it as tracing "only models whose training data is open" (census record); Farhadi ties the connection to Ai2's open inputs, and Liu expects closed-model providers to hesitate because the tool would expose their training data (Ken Yeung). Its purpose, in the paper, is to explore where models "may have learned to generate certain word sequences", verbatim matching taken as "the most direct connection between LM outputs and the training data" (the paper), for researchers, developers and the general public (Ai2 blog), at "an early stage" of exploring the tool (Ai2 blog). The archive places the class this tool belongs to on the training/output relation: data attribution asks "Which training examples shaped a specific output", an adjacent method that "Operates on training/output relation; PER operates on output/reader relation" (#725). Of that class the archive holds that partial tracing "is possible at cost", that it does not recover authorship, and that "what survives depends on what is recorded and computed" (#1714), and that training-data attribution is the post-hoc detection layer, developed "in isolation" from the other survival layers, no system bundling them (#127). Claims drawn from a model's training-time parameters, on the archive's typology, often carry no surface signal of the sources that contributed; PER reaches them only indirectly, and the substrate level needs companion infrastructure, "training-data provenance records" among it (#95), since provenance "has to survive the training pipeline" (#1714). Its authors show it on "fact checking, hallucination, and the creativity of language models" (arXiv abstract), with three case studies, fact checking first (the paper), and Ai2's post traces a math expression the model could have learned by memorization and a false knowledge cutoff to post-training examples, a finding that changed the 32B model's post-training data (Ai2 blog). Fact checking stands in the paper as a possibility, users "may be able to fact-check the statement against its training data", with its qualification that context and source help "a more informed judgment on the factuality of the statement", since words mislead out of context and some web sources are unreliable (the paper, arXiv v2 and ACL 2025). Ai2's post calls the interpretation of matches "quite nuanced" and makes confidence in a generated fact conditional on the user's own check that the documents "come from trusted sources" (Ai2 blog); the newsletter calls it an application "offering fact-checking", and Liu's launch statement says users "can ensure that the data supporting their outputs is trustworthy and verifiable" (Ken Yeung). The authors state the limit in both versions of the paper, "The retrieved documents should not be interpreted as having a causal effect on the LM output, or as supporting evidence or citations for the LM output" (the paper; ACL 2025), and the census carries it as two negative tags, not causes of the output and "NOT a citation system", and in what it records as surviving (census record). To the newsletter Ai2 called the documents associations, not "the definitive source of the model's learning", and its results not "indicators of authority" (Ken Yeung). Ai2's leaders tie the tool to hallucination scores, "assuming that the inputs were factual" (Farhadi), and to regulation, governance and auditability (Lebrecht), in the newsletter's report (Ken Yeung). The paper sets it apart from retrieval-augmented generation, naming Google AI Overview among RAG's examples: OLMoTrace "retrieves documents post-hoc and does not intervene with the LM generation" (the paper; Liu in Ken Yeung), and Gemini's double-check, drawn from Search and not from Gemini's training data, is "less useful for scientific exploration" (the paper). At its own name on Google's AI Overview, 2026-10-29, the composition held the tool, its maker, its real time, its scale and its verbatim matching, checked against the census record and the pages it saved, naming the tool, its method, its models and its uses (Capture Registry, 2026-10-29), and gave as its first use "Fact-Checking: Verify whether specific claims in a model's answer are backed by real training documents", a verbatim match "offered as support for a claim, the reading the paper excludes" (Capture Registry, 2026-10-29). The registry reads this as "the conversion of a matching relation into a citation relation that the paper names and refuses", a tool whose authors state what a match does not mean "composed with that meaning restored", and it finds the census record's negative tags to be the authors' own (Capture Registry, 2026-10-29). Its 31 authors are at the Allen Institute for AI, the University of Washington, UC Berkeley and Stanford (arXiv; the paper), the operator the census records whole (census record), where the composition named Ai2 alone (Capture Registry, 2026-10-29). It is "publicly available and fully open-source" (arXiv abstract), its core under the Apache 2.0 license (the paper), reached through the Ai2 Playground (census record), an expression of Ai2's commitment to "open models, open data, and beyond" (Ai2 blog). Submitted 9 April 2025 and revised 8 July (arXiv; census record), it appeared in ACL 2025's System Demonstrations with the Best Demo award (ACL Anthology), and the census records it active, "ACL 2025 demo track" (census record). Its authors name as a risk that it makes problematic training content "more easily exposed" (the paper), and the newsletter finds that it gives no analytics for improving the data: "Simply put, it's generating awareness" (Ken Yeung). The census records the tool as one that records provenance, a neighbour of the Data Provenance Initiative, with the Playground and "infini-gram as a separate record" left unchecked (census record); no card cites the record, which landed the same day (Capture Registry, 2026-10-29), and Wikipedia has no article on OLMoTrace (Wikipedia, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'OLMoTrace' (T1, six cards: B1–B6, B4 and B5 unreachable; 207 words) the composition gives the tool by its authors' abstract and launch post. Composed from its cards: the definition at the abstract's and the post's grain (B2 F1, B1 F2: tracing outputs back to multi-trillion-token training data in real time; 'full' dropped), Ai2 as maker (B1) with the partners left out (B2 F43 names 31 authors; the affiliations are on B7 F42, the paper the cards link to: CAP-1-05); open source (B2 F45); the button and the verbatim spans (B5 F12, snippet; B1 F4); the index and its complexity (B1 F9); up to 10 documents ranked by BM25 (B1, the post's text of the pipeline; B7 F10); the models (B1 F17, B4 F18 for OLMo 3, the X snippet, CAP-1-06); the uses at the abstract's and the newsletter's grain (B2 F21 'fact checking, hallucination'; B6 F28 'offering fact-checking'; B1 F31 memorization; B6 F33 regulation and auditability). Force raised: the fact-checking use. The field states it as a possibility, 'users may be able to fact-check the statement against its training data' (B3 F24, the card's own paper; B7 F23), qualified by context and source (B7 F25), conditioned in the operator's post on the user personally verifying that the documents 'come from trusted sources' (B1 F27) and called 'quite nuanced' (B1 F26); T composes 'Verify whether specific claims in a model's answer are backed by real training documents', an imperative of support. Limit dropped: the authors' limit, in the card's own paper, that the documents are not 'supporting evidence or citations for the LM output' (B3 F35; B7 F34), and Ai2's statements to the newsletter that they are associations and not 'indicators of authority' (B6 F36, F37): none is composed. The strongest form of the use in the field is Liu's launch statement, users 'can ensure that the data supporting their outputs is trustworthy and verifiable' (B6 F29, attributed); T's 'backed by' stands nearer to it than to the paper, at no attribution. The capture's findings name the same move at their own grain: 'A verbatim match is offered as support for a claim, the reading the paper excludes' (CAP-1-04), 'the conversion of a matching relation into a citation relation that the paper names and refuses' (CAP-1-09). Read against the field whole, fact checking is the authors' own first use (B2 F21; B7 F22) and what the paper excludes is reading the matched documents as causes, evidence or citations: the excess of T is the support relation and the dropped limit, not the use. Substituted reading: T's 'ranks up to 10 context documents using BM25 and unigram probabilities' joins two steps the field keeps apart (the span unigram probability filters spans, B1 'OLMoTrace behind the scenes' ¶4; BM25 ranks documents, B7 F10). Outside the cards: 'such as the Dolma datasets' (no field source names Dolma; the paper's own page, not surfaced, names olmo-mix-1124 and dolmino-mix-1124); 'or generated from general parameters' (no field source makes the dichotomy; the archive's typology of parametric claims, O95-01, bears on it); 'mixes up facts' and 'data recipes and pretraining paths' (no field source). Available, not composed: the count (B1 F16), the purpose (B1 F7), the early stage (B1 F8), the hedge on relevance (B1 F11), the condition of access and the closed providers (B1 F17 in part; B6 F20, F47), the knowledge-cutoff finding (B1 F32), Farhadi's conditional (B6 F30), the contrast with retrieval (B6 F40), the openness as commitment (B1 F44), the publication and its Best Demo award (B2 F48, B3 F49), the newsletter's assessment (B6 F38). Composed without attribution: everything; no person is named, at the coarse grain of a tool description (recorded, not a breach), except the use, which is composed at the grain of a support relation no card states for the documents (recorded as force raised). Content replaced by use: the closer (T1-9).

v0.9 §0.5: the first diff is ontological and narrow. T and the field agree on the entity (Ai2's tool that finds training text verbatim in a model's answer). T composes the tool and its uses and leaves out the one sentence its authors attach to what the tool returns: that a match is not a cause, not evidence, not a citation. With the limit gone, the matching relation is composed as a support relation.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the class's relation, training/output and not output/reader, where PER operates (O725-01); partial tracing possible at cost, recovering no authorship, what survives depending on what is recorded and computed (O1714-01), provenance that has to survive the training pipeline (O1714-02); training-data attribution as the post-hoc detection layer, developed in isolation from the other survival layers (O127-01); the archive's evidentiary hierarchy for training-data attribution, with unvalidated similarity as exploratory evidence only (O1075-01, O1075-02), set beside the paper's heuristic of lexical overlap (F5); parametric claims, which carry no surface signal of their sources and need training-data provenance records at the substrate level (O95-01); the provenance census record's answer to its organizing question, field by field with its basis (PE-01–PE-21), its tag 'traces only models whose training data is open' held against the field's condition of access (PE-17 against F17, F19); and the Capture Registry's account of the composition at the name (CAP-1-01–CAP-1-11). Not surfaced: every archive source; no card cites the census record (CAP-1-11).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F7', 'F8', 'F9', 'F11', 'F12', 'F16', 'F17', 'F18', 'F20', 'F21', 'F24', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F35', 'F36', 'F37', 'F38', 'F40', 'F43', 'F44', 'F45', 'F47', 'F48', 'F49'], 'T1_composed': ['F1', 'F2', 'F3', 'F4', 'F9', 'F10', 'F12', 'F17', 'F18', 'F21', 'F24', 'F28', 'F31', 'F33', 'F45'], 'T1_outside_cards': ["'such as the Dolma datasets' (no field source; B7, own page not surfaced, names olmo-mix-1124 and dolmino-mix-1124)", "'or generated from general parameters' (no field source)", "'mixes up facts' (no field source)", "'data recipes and pretraining paths' (no field source)"], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}