mjbommar/openlemma-v0.4-pretrain

OpenLemma v0.4: pretraining documents (v0.4.1)

Two families of verified procedural and reasoning documents from OpenLemma v0.4. Every document here is one the ALEA R3 model trains on in its stable phase, token for token, with the same filters. Everything is generated and checked by code; no language model wrote any of it.

config	documents	ALEA tokens	passes	share of the stable mix	what
ol_single	1,482,879	91,954,045	2.86	0.80%	one verified exercise per document, ending on its answer: 31 families (analogies, fractions, units, dates, multi-step arithmetic, percentages, money, letter counting, sequences, truth tables, set operations, syllogisms, conditional inference, string and list rules, bit-vector equations, constraint puzzles, format conversion and repair, JSON patches, schema checks, code tracing); 9.5% with worked steps, 11.9% in chat form
demap_pe	99,705	12,159,636	4.05	0.15%	one few-shot episode per document: extraction (money amounts, dates, citations, CFR references, docket numbers, day counts, percentages; 49% of documents), procedural string and number rules, and label tasks, in ten pair layouts

Tokens: alea-p1sw-staged-16k, including <|start|>/<|end|>, 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: the openlemma-v0.3-pretrain schema (id, text, source, family, group, tier, k, n_tokens, audit_accuracy, families, groups, item_ids, n_tokens_alea_p1sw, origin, proc_v1_id) plus meta (JSON: the item with its answer, layout, format and whether worked steps are shown). n_tokens and n_tokens_alea_p1sw both hold the ALEA token count.

The OpenLemma v0.1-v0.3 documents in the same stable mix come from mjbommar/openlemma-v0.1-pretrain, -v0.2-pretrain and -v0.3-pretrain (the mix trains filtered subsets of v0.1 and v0.2 and a re-render of v0.3).

CHANGES (v0.4 -> v0.4.1): 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: 918 documents that share an n-gram with ALEA evaluation items added since v0.4 (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 v0.4 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 v0.4	rows v0.4.1	trained in stable.v4d	in the set, unplanned	kept, not in the mix	exposures per trained document in stable.v4d
demap_pe	99,834	99,705	99,705	0	0	1.66
ol_single	1,483,668	1,482,879	1,482,879	0	0	1.17

How it was built and checked

• ol_single: each item is built by construction and re-solved by its family's own solver and checker; worked steps are recomputed with exact fractions, including the final step; every document ends on the answer (0 exceptions), so the model learns to stop after answering instead of starting a new exercise. Question framings are capped so that none is in more than 1% of documents; fraction questions are re-framed; real-word propositions come from OpenGloss.
• demap_pe: every pair is re-derived (the generating function, an independent regex parser for extraction, the label map), and every document ends on the answer to its final query. Pair layouts include x -> y, x: y, tab, |, =, Input: / Output:.
• Decontamination (fail closed): the R3 decontamination bundle (SHELF validation/test 13-grams and every ALEA evaluation), template-leak 8-grams, copyright-evaluation 13-gram hashes, and the later evaluation references (ICL nonce words and pairs, held-out relation pairs, and the operand pairs of ALEA's arithmetic-across-notations probe: 5,770 ol_single and 43 demap_pe documents removed).
• Toxicity: every document was scored with s-nlp/roberta_toxicity_classifier; documents with any 64-word chunk

  = 0.8 were removed (ol_single 1,046, demap_pe 214).

Held out and not released

• A separate 40% split of each family (a seeded document split) is not in the stable mix and is not released here.
• ALEA's arithmetic-across-notations probe and other evaluation items are not released.

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
ol_single	1,482,879	0 / 0 / 0	1 / 1 / 414	8 / 21	409	every document scored; >= 0.8 removed	0
demap_pe	99,705	0 / 0 / 0	0 / 0 / 0	719 / 614	216	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 'phone' hits are number sequences in pattern exercises that the phone pattern misreads; the street hits are invented business names in extraction sentences ("Elm Street Bakery ...") and titles in lists. All names and records are invented; no real person's contact details 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, held-out pairs or probe operands. A pair-level check flags 41 documents whose pairs also appear as demonstration pairs of ALEA's held-out relation probe (trained by design); none is a held-out query pair.

Provenance and licence

• Generation and release code will be published with the ALEA model.
• Sources: OpenLemma generators (the author's own, CC-BY-4.0); OpenGloss vocabulary (CC-BY-4.0; WordNet notice in LICENSE-WordNet.txt); invented names and records; short U.S. government text snippets (public domain) as extraction payloads. Released under CC-BY-4.0.