mjbommar/openlemma-v0.3-eval

OpenLemma v0.3: evaluation (ALEA proc-v1 families)

Held-out items, never in the training data (split test). They were generated fresh with proc-v1's generators under new seeds (openlemma-v0.3-eval:<family>:<n>), and every item whose problem key occurs anywhere among proc-v1's 48.6M generated items (released or not), or whose content matches an OpenLemma v0.1/v0.2 item, was skipped. Decision labels are balanced per group. Training documents sharing a 13-gram with an eval item or probe were removed.

• item configs (default, one per group): input holds sub, problem and question (the rendered question text); answer is canonical; record.answer_text is the rendered answer;
• probes: copy-rate probes in the v0.1 format. input is a k-shot block (or an induction episode) cut just before the query's answer; options has four distinct strings: the answer, the copy option (the previous item's answer, or the last demonstration's output), and two answers of other held-out items of the family. record has answer_index, copy_index, kind, k, layout, item_ids. Score each option as a continuation of input; report accuracy and copy rate (chance 25%). Decision families have no probes (their two options would be the labels).

Config	Rows
default	15,500
pattern	5,000
logic	2,500
arith	3,000
formats	3,500
code	1,500
probes	2,700

Items per family: proc.analogy 500, proc.arith_multi 500, proc.bitvec 500, proc.code_equiv 500, proc.code_trace 500, proc.cond_inference 500, proc.constraint_puzzle 500, proc.copy_modify 500, proc.count_pos 500, proc.csv_edge 500, proc.dates 500, proc.fmt_convert 500, proc.fmt_query 500, proc.fmt_repair 500, proc.fractions 500, proc.json_patch 500, proc.kv_recall 500, proc.list_induce 500, proc.list_ops 500, proc.map_induce 500, proc.money 500, proc.num_induce 500, proc.percent 500, proc.schema_infer 500, proc.schema_validate 500, proc.seq_numeric 500, proc.seq_symbolic 500, proc.set_ops 500, proc.syllogism 500, proc.truth_table 500, proc.units 500.

Provenance

This is ALEA's procedural dataset proc-v1 (ALEA notes/38; code in the ALEA repository, experiments/procedural), published as part of OpenLemma as v0.3. Release notes: docs/RELEASE-V0.3.md. It is a new release beside v0.1 and v0.2, which are unchanged.

• Deterministic, $0. Every problem is generated by construction from a seeded generator (seed proc-v1:<shard>:<doc>); no language model wrote any row, and no API was called.
• Every item is verified, fail-closed. Each family recomputes the answer from the stored problem by a different code path (rule-hypothesis enumeration for induction, truth tables / model enumeration for logic, exact rationals, datetime vs Rata Die, standard parsers and jsonschema / jsonpatch / jsonpath-ng / sqlite3 for formats, restricted execution and exhaustive search for code). For v0.3 every item of every released document was re-checked in freshly spawned processes: 21,437,761 of 21,437,761 pass; the rendered answer is a faithful rendering of the canonical one and occurs in the document.
• Negative controls: 2,141,910 of 2,141,910 corrupted answers (last character changed, JSON/code broken) and 1,905,039 of 1,905,039 swapped answers (another item's different answer of the same sub-task), and 328,002 of 328,002 flipped decision labels (valid/invalid, yes/no, T/F ...) were rejected by the same checkers.
• Text is byte-identical to proc-v1. v0.3 only removes documents; proc_v1_id and mapping/proc_v1_documents.parquet (pretrain repository) map every proc-v1 document to its v0.3 id or the reason it was dropped. ALEA's R2 decay (v3) trained on a 1,006,966-document sample of proc-v1; 915,403 of those documents are in v0.3 (flag in_alea_r2_decay_v3).
• Item ids are content ids: proc.<family>:<sha256(family NUL problem JSON)[:16]>, so the id of any proc-v1 item can be recomputed from proc-v1's raw shards.
• Deduplicated against OpenLemma v0.1 and v0.2 (pretrain documents, items and evals): exact documents, exact item content, and MinHash near-duplicates for the formats and code families.
• Decontaminated (fail-closed) against ALEA's evaluations (ICL v1/v2, OOD-ICL, knowledge probe, general-facts probe, likelihood suite, judge prompts, chat probe, template-leak prompts, QA rounds, structured-formats baseline, closed-book facts), SHELF validation and test, and the OpenLemma v0.1, v0.2 and v0.3 evaluation sets.

Token counts use ALEA's alea-p1-prefix-16k (as for v0.1 and v0.2); n_tokens_alea_p1sw gives alea-p1sw-staged-16k counts (without <|start|>/<|end|>), the tokenizer of ALEA's proc-v1 build.

Licence and attribution

CC-BY-4.0. Attribute OpenLemma (https://github.com/mjbommar/openlemma) and ALEA (proc-v1). Values inside items come from:

• Wikidata (2025-11 dump; CC0): entity labels, descriptions and facts in formats items, names, analogies;
• OpenGloss v2.4 (CC-BY-4.0): headwords and short glosses. OpenGloss v2.4 includes lexemes derived from Princeton WordNet 3.0; such content carries the WordNet licence, included as LICENSE-WordNet.txt.

Known limitations

• Families are procedural: breadth comes from many sub-tasks, value pools and surfaces, not from free text. code_trace uses 14 program templates and code_equiv 8 function shapes (identifiers renamed).
• Documents mix 1-4 blocks from different families, so a group config holds only the single-group documents; filter default on groups for every document that contains a group.
• No language-model judgement is involved anywhere; nothing here is agreed-tier.
• ALEA's R2 decay saw part of this data before publication (above); no claim about downstream gains.

Citation

@misc{openlemma2026,
  title  = {OpenLemma: verified, bottom-up synthetic data for logic, reasoning, algorithms and language},
  author = {Bommarito, Michael J.},
  year   = {2026},
  note   = {Version 0.3: ALEA procedural data (proc-v1)},
  url    = {https://github.com/mjbommar/openlemma}
}