OpenGloss-ICL v0.2

Deterministic in-context-learning documents built from verified OpenGloss v2.4 word relations and the OpenGloss-Facts Wikidata knowledge base. Every document here is one the ALEA R3 model trains on in its stable phase, token for token, with the same filters. No language model wrote any of this text.

config	documents	ALEA tokens	passes	share of the stable mix	what
relations_gold (default)	490,092	45,278,202	4.36	0.60%	WordNet-verified relation pairs (antonym, synonym, hypernym, inflections, comparative/superlative) as k-shot lists, relation classification with natural, flipped or invented labels, short Q/A, chat turns and multi-episode documents
word_map	293,063	42,204,505	4.29	0.55%	one mapping rule per document over the same gold pairs (relations and morphology), half of them bare 4-12-pair blocks whose rule must be inferred from the examples
arbitrary_label_icl	114,086	22,799,678	4.29	0.30%	single-word classification with arbitrary labels defined only by the demonstrations (part of speech, category, entity type, continent, number and word shape)

Tokens: alea-p1sw-staged-16k, including <|start|>/<|end|>, exactly as trained. Shares are of the stable.v4c mix used before 2026-10-12 (32,882,670,081 tokens including passes). mapping/stable_v4c.parquet gives every document's id, text SHA-1, config, training set, passes (exposures) and token count.

Columns: id, text (the training document), family, sub (task), format, layout, messages (chat documents only), meta (JSON: the pairs or classes used, with their relation), provenance (JSON: source kinds), generator, n_tokens_alea.

CHANGES (v0.1 -> v0.2): 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: 118 documents that share an n-gram with ALEA evaluation items added since v0.1 (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 v0.1 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 v0.1	rows v0.2	trained in stable.v4d	in the set, unplanned	kept, not in the mix	exposures per trained document in stable.v4d
arbitrary_label_icl	114,204	114,086	114,086	0	0	1.76
relations_gold	490,092	490,092	490,092	0	0	1.78
word_map	293,063	293,063	293,063	0	0	1.76

How it was built

• Gold pool. OpenGloss v2.4 relation edges on a word's primary sense, kept only where WordNet 3.0 agrees (76,663 gold pairs; a small silver set of antonym pairs is used only in plain statements). Evaluation words, nonce strings, generic hypernym targets and gerund plurals are removed. Each pair is used at most 250 times across relations_gold and word_map, with Zipf-weighted draws so the tail of the vocabulary is used.
• Layouts. 60.5% of relations_gold documents use compact mapping layouts (x -> y, x: y, tab-separated, x = y, Input: / Output:), so these layouts carry meaning-based relations and not only string edits; 14.1% of relation documents use flipped or invented labels, so the label has to be read from the demonstrations. => is never used.
• One rule per document in word_map (no document starts a second block after its answer); in relations_gold 11.6% of blocks are followed by another block.
• Cue caps. Every framing string and opener is in at most 0.3% (relations_gold) or 0.95% (word_map, arbitrary_label_icl) of documents.
• Decontamination (fail closed). The R3 decontamination bundle (including SHELF validation/test 13-grams and every ALEA evaluation), template-leak 8-grams, copyright-evaluation 13-gram hashes, and the later in-context-learning evaluation references (mapped evaluation words, nonce strings, held-out relation pairs, arithmetic-probe operands). arbitrary_label_icl lost 5,342 documents to these references.
• Toxicity. Every document was scored with s-nlp/roberta_toxicity_classifier; documents with any 64-word chunk

  = 0.8 were removed (relations_gold 6,686, word_map 8,315, arbitrary_label_icl 144). Bare word lists still contain ordinary dictionary words that are vulgar or sensitive out of context (counts below).

Held out and not released

• 250 held-out antonym pairs and 896 held-out pairs of the other relations, used by ALEA's held-out relation probe: no document maps a word of a held-out antonym pair, and no document puts both words of a held-out pair on one line.
• A separate 40% split of word_map and arbitrary_label_icl (a seeded document split) is not in the stable mix and is not released here. relations_gold trains in the stable mix in full.

Pre-release screens

config	docs	e-mail (official / placeholder / other)	phone (official / placeholder / other)	street (official / other)	hashed slur/vulgar lexicon (docs)	RoBERTa toxicity >= 0.8	decon hits
relations_gold	490,092	0 / 0 / 0	0 / 0 / 0	0 / 0	2,880	every document scored; >= 0.8 removed	0
word_map	293,063	0 / 0 / 0	0 / 0 / 0	0 / 0	4,661	every document scored; >= 0.8 removed	0
arbitrary_label_icl	114,086	0 / 0 / 0	0 / 0 / 0	0 / 20	844	every document scored; >= 0.8 removed	0

• PII: private contact details only. official = government domains, or a number/address next to an agency, court or office cue; placeholder = example domains, 555 numbers. No text was changed: 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; the classifier filter above was applied to every document.
• PII review: the 20 'other' street hits are titles and names in labelled lists (for example a film called Wall Street) that the address pattern misreads. No contact details of any person were found.
• Decontamination re-check on the release files: 0 hits for the R3 bundle (including SHELF validation/test 13-grams), the template-leak 8-grams and the copyright-evaluation 13-gram hashes; 0 evaluation nonce strings and 0 held-out pairs on one line. A pair-level check flags 189,663 documents whose pairs also appear as demonstration pairs of ALEA's held-out relation probe; those demonstration pairs are trained by design (the probe compares seen and held-out pairs), and none of the hits is a held-out query pair.

Known limitations

• Relation pairs are verified against WordNet but some remain loose (an antonym list can contain a weak contrast) and a few morphology forms are rare or unusual (plurals of mass nouns).
• Lists are context-free: a sensitive or vulgar headword can appear next to ordinary words.

Provenance and licence

• Generation and release code will be published with the ALEA model.
• Sources: OpenGloss v2.4 (the author's own dataset, CC-BY-4.0; WordNet notice in LICENSE-WordNet.txt) and Wikidata (CC0) through the OpenGloss-Facts knowledge base. Released under CC-BY-4.0.

Citation

@misc{bommarito2025opengloss,
  title  = {OpenGloss: A Synthetic Encyclopedic Dictionary and Semantic Knowledge Graph},
  author = {Bommarito, Michael J., II},
  year   = {2025},
  eprint = {2511.18622},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2511.18622}
}