OpenGloss-Facts (v1.3)

Verified, multi-form world knowledge from Wikidata (CC0) for small language models. Version 1.3 is exactly the OpenGloss-Facts data that the ALEA R3 model trains on in its stable phase (mixes stable.v3 to stable.v4c, steps 5,800-23,200; from step 23,200 the stable.v4d mix, see CHANGES), token for token.

config	documents	ALEA tokens	passes in stable.v3	share of stable.v3	what
deterministic	4,939,810	365,279,448	2.836	3.210%	v1.1 KB and frames re-rendered one subject per document (render v3)
gpt6luna	261,757	15,239,515	2.0	0.095%	model-generated (gpt-6-luna): the single-subject subset of the v1.1 texts
canon (new)	349,625	13,167,432	10.0	0.408%	everyday-canon head facts (315,917) and event timelines (34,032); deterministic
eval	2,032	n/a	not trained	—	closed-book recall eval v1, byte-identical to v1.1 (canary in every row)

Tokens: alea-p1sw-staged-16k incl. <|start|>/<|end|>, as trained. mapping/stable_v3.parquet: id, config, training set, passes.

CHANGES (v1.2 -> v1.3): tracking the stable.v4d mix

On 2026-10-12 (step 23,200) ALEA R3's stable phase switched to a new mix, stable.v4d (13.5B tokens; it trains until the stable phase ends at step 35,619). It removes a few layouts the model had over-learned, cuts the correction share, applies the latest evaluation references, and adds the NLP task suite. This version tracks it:

• Removed from this version: 7,664 documents that share an n-gram with ALEA evaluation items added since v1.2 (held-out task items, probe prompts), outside the boilerplate exemption for public-record text. stable.v4d's fail-closed decontamination removed the same documents from training. They remain in the v1.2 tag.
• New file mapping/stable_v4d.parquet: one row per released document (and the few stable.v4d documents that are not released): id, text_sha1, config, stable_set, v4d_status, exposures (passes in stable.v4d), n_tokens_alea (as trained, incl. <|start|>/<|end|>), released. v4d_status is trained, kept_unplanned (in the set, but after the point where its planned fraction of a pass ends), removed_format, removed_decon / removed_decon_screen (not released), removed_tox (not released) or not_in_v4d.
• Every other file is byte-identical to the previous version; configs with removed documents keep their shard names, schema and row order.

config	rows v1.2	rows v1.3	trained in stable.v4d	in the set, unplanned	kept, not in the mix	exposures per trained document in stable.v4d
canon	349,949	349,625	349,625	0	0	4.10
deterministic	4,946,299	4,939,810	4,939,810	0	0	1.16
gpt6luna	262,608	261,757	215,067	46,690	0	1.00

CHANGES (v1.1 -> v1.2)

• deterministic: re-rendered for one subject per document (same kb_v1 train split, same relation frames and independent checker). Ids are new (ofv3-*); family is the document family. Families: profile 1,705,369, qa_plain 922,784, chat_qa 802,593, reverse_index 573,882, fact_list 348,223, table 272,573, cloze 201,401, contrast 91,949, timeline 27,525.
  • Frame caps: every sentence/question/heading/system-prompt frame is capped per document at <= 0.3% (max 0.279%, 5,639 frames); "whose nationals included" 1.06% -> 0.27%; multi-subject documents 67% -> 0.
  • Sign-offs: both sign-off lines removed (0.70% / 0.74% -> 0).
  • Q/A shape: Q/A documents are one pair ending the document in 67.8% of cases, the rest 2-3 pairs about the same entity; 0 documents with >= 4 Question: labels; every question names its subject in full.
  • Dropped families: companions, multi-entity tables, quizzes, compare/chain/glossary; reverse_index is answer-first; table is a single-entity fact box. About 1.3% are single-turn abstention chats about an invented name.
  • Redundant facts: 2,182 documents with redundant "The {N} of {S} is in {O}." facts ("the river of the Susquehanna River") dropped at assembly.
  • Fail-closed checks at render: the v1.1 checker, decon (12,087 dropped), the general-knowledge eval guard (92).
• gpt6luna: the single-subject subset : 419,886 -> 262,608. Documents mixing several subjects (37%) are dropped, openings capped at 0.9%; 5 removed by the general-knowledge eval guard. No new generation; rows are byte-identical to v1.1.
• canon: new config. Head entities of the everyday canon (authors/works, painters, composers, scientists, heads of state, capitals, landmarks, inventions, elements, geography): 6,994 entities, 37,880 facts, each re-derived from the raw Wikidata claims (facts that do not re-derive are dropped; vandalised labels filtered), rendered 2-20 times in different surfaces, single subject, frames <= 0.40%, openings <= 0.9%. Events: phases, actors and dates of historical events from the same dump. Event rows are decoded from the trained token sequences (the part file on disk was re-rendered after training data was built) and accepted only on an exact token round trip (34,032 / 34,032).
• eval: unchanged (same 2,032 items, ids and canary).

Known issue, left as trained: some canon/event frames read stiffly (" is found at a location in ", " was when the end of occurred"); the next canon render rephrases them.

The gpt6luna config is model-generated

Natural paragraphs, dialogues, letters and narratives written by OpenAI gpt-6-luna (flex, low effort; $39.30 for the v1 run) around Wikidata fact bundles; every value, name and number checked by code against the facts (v1.1 card).

Verification

• deterministic (v3 render): 100% pass the v1.1 checker in fresh processes; frames <= 0.3%; round trip 3,000 / 3,000.
• canon: the v1.1 checker against canon_kb; 99.99% single-pair Q/A; frames <= 0.40%.
• Release build: every row hashes to its trained document (SHA-1 of text and token sequence); facts re-read from the KB.

Decontamination and held-out data

All train configs passed R3's fail-closed decontamination and the late template-leak and copyright-eval scans. The held-out entities of the eval config are absent from the train configs. Facts data used only in the final training phase (e.g. the canon gpt-6-luna paraphrases, made-up-entity abstention) is not included and stays private until the model ships.

Pre-release screens

config	docs	e-mail (official / placeholder / other)	phone (official / placeholder / other)	street (official / other)	hashed slur/vulgar lexicon	RoBERTa >= 0.8 (sample)	decon hits
canon	349,625	0 / 0 / 0	0 / 0 / 0	0 / 6	39	1/2000 (est. 175 docs)	0
deterministic	4,939,810	0 / 0 / 0	0 / 0 / 0	1 / 1	1,927	1/2000 (est. 2,473 docs)	0
gpt6luna	261,757	0 / 0 / 0	0 / 0 / 0	0 / 0	222	0/2000 (est. 0 docs)	0

• PII: private contact details only; names in the public record (parties, judges, officials, statute names) are never screened or removed. official = government domains, or a number/address next to an agency, court or office cue; placeholder = example domains, 555 numbers. 'Other' hits are counted and listed by id for review; no text was changed, because this release mirrors the trained documents exactly.
• Toxicity: the hashed slur/vulgar lexicon of the ALEA content guard (no terms stored in clear) on every document, and the KL3M behaviour suite's classifier (s-nlp/roberta_toxicity_classifier, any 64-word chunk >= 0.8) on a seeded sample per config. No document was removed: stable training applied no toxicity filter, and this release tracks it.
• PII review: The street hits are landmark and building names from Wikidata labels.
• Decontamination re-check: 0 hits for the R3 bundle (including SHELF validation/test 13-grams), the post-bundle template-leak 8-grams and the copyright-eval 13-gram hashes.

Provenance and licence

• Generation, checking and release code will be published with the ALEA model.
• Wikidata (2025-11 dump, local copy): CC0 1.0. Generated text: CC-BY-4.0. gpt-6-luna outputs: generated by OpenAI gpt-6-luna, released under CC-BY-4.0, subject to the provider's terms for outputs. No Wikipedia, OpenGloss or WordNet text.

v1.1 remains at the tag v1.1.

Citation

Bommarito, M. (2026). OpenGloss-Facts: verified, multi-form world knowledge for small language models.