mjbommar/openlemma-v0.2-eval

OpenLemma v0.2: evaluation

Held-out items, never in the training data (split test):

• code: every code surface on held-out seeds; kernel-based surfaces use 12 kernels that never appear in training;
• code_llm: the 259 verified exercises of the gpt-6-luna pilot, for sandboxed pass@1 (run the tests in input against a model's function);
• formats: every formats family on held-out seeds.

Training documents whose content matches a held-out item were removed.

Config	Rows
code	4,254
code_llm	259
formats	2,365

Provenance and verification

Built by OpenLemma (code: MIT) from the plan in docs/FORMATS-CODE-V2-PLAN.md; release notes in docs/RELEASE-V0.2.md. This is a new release beside v0.1, which is unchanged.

• Generation: deterministic procedural generators (seeded; no randomness at build time) and, for code2-llm only, gpt-6-luna (OpenAI, flex tier, reasoning effort low). No other model wrote any row.
• Every row is verified (fail-closed: anything unchecked or failing was dropped):
  • code: every program ran in a sandbox (bubblewrap: no network, read-only system, private /tmp, CPU/memory/time limits, an audit hook refusing sockets, subprocesses and writes) and was compared, on the host, with an independent oracle: a brute-force implementation written differently, gcc for C, exhaustive search over a bounded domain, recompiled bytecode, finite differences or exact rationals. Test suites must fail on most single-point mutants;
  • formats: each answer is re-parsed or recomputed with an independent library (json, PyYAML, tomllib, ElementTree, jsonschema, jsonpatch, jsonpath-ng, sqlite3, csv).
• Negative controls: for a sample of items a wrong answer (a mutant, a flipped verdict, an unfixed bug, a changed value) was run through the same checker; the rejection rates are in manifest.json.
• Anti-template rules: sentences are composed from many alternatives, code-block and heading styles rotate, and a monitor over the finished text keeps every opening line, label and prose line at or below 1% of documents. Surfaces used by ALEA's evaluations are never emitted.
• k-shot: formats documents hold 1-3 items (60/30/10%) of one family, with pairwise-distinct answers; labelled items are balanced. Copy-last and majority-label baselines are reported per family.
• Decontamination: documents were checked against ALEA's evaluation sets (knowledge probe, ICL v1/v2, OOD-ICL, likelihood suite, general-facts probe, judge prompts, structured-formats baseline, chat probe, QA rounds), SHELF validation and test, OpenLemma v0.1's eval release, and this release's held-out items; overlapping documents were removed.

Token counts use the ALEA tokenizer alea-p1-prefix-16k (as in v0.1).

Licence and attribution

CC-BY-4.0. Attribute OpenLemma (https://github.com/mjbommar/openlemma). The value_pool rows and the values inside formats items come from Wikidata (2025-11 dump), which is CC0. No WordNet-derived material is included. gpt-6-luna output is subject to the provider's terms.

Known limitations

• Kernel-based code surfaces reuse a 118-function library (with renames and varied text); synthesized surfaces and the LLM exercises supply the breadth. 12 kernels are held out for evaluation only.
• decompile listings are pinned to CPython 3.14; other versions print different bytecode.
• Decontamination against HumanEval/MBPP/APPS was not run (no local copy); the code is generated, not scraped.
• Published before any model was trained on it; 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 0.2: formats v2 and verified code v2},
  url    = {https://github.com/mjbommar/openlemma}
}