LuzCommand/lil-beastie-4b-eval

lil-beastie-4b — the complete evaluation bundle

Everything needed to reproduce every number on the lil-beastie-4b model card from raw data, or to re-run the core evaluation from scratch for ~$2 of cloud GPU.

Reproduce the card WITHOUT a GPU (from the shipped raw rows)

MMLU-Pro calibration (multiple choice, vote-based):

python harness/score_g10.py --run results/ck23_g10_fresh2 --exclude 4849,10565 --boot 300

→ ECE 0.082 · Brier 0.162 · AUROC 0.780 · answered 94.0% · precision 75.7% · the full consensus curve · per-category table. Same scorer reproduces every other MMLU-Pro run in results/.

Free-form AIME on the 58 never-trained problems (the 38 training-lineage problems excluded):

python harness/score_freeform.py results/aime_ab_brief_q8 results/aime_ab_through_q8 \
    --exclude-file contamination/aime_training_lineage_ids.json

→ commit precision 97.8% (220/225) "Reason briefly" · 98.9% (182/184) "Reason it through" · paired −0.66 correct per 10, t = −3.54.

HumanEval, lil-beastie vs the stock base on the identical harness:

python harness/score_humaneval.py results/coding_bench_g5/results.jsonl \
    --base results/coding_bench_g5_base/results.jsonl

→ pass@1 92.9% vs 92.8% · pass@5 96.3% vs 97.6% · paired t = 0.12 · 22% less generation time.

Side benches (lil-beastie vs the base and Qwen3-4B-Instruct-2507, G=5): PopQA, SelfAware, SQuAD 2.0, GSM8K/MATH-500 and IFEval:

python harness/score_popqa.py decks/side_benches/popqa_1k.jsonl results/side_benches/popqa_{lilbeastie,base,instruct}
python harness/score_qa.py decks/side_benches/squad2_1k.jsonl results/side_benches/squad2_{lilbeastie,base,instruct}

→ PopQA confidently wrong 2.9% / 23.8% / 44.5% · SQuAD 2.0 abstention F1 76.4 / 65.2 / 72.7. The full command list, setup and caveats are in results/side_benches/SUMMARY.md. harness/verify_traces.py <deck> <run_dir> checks every graded prediction against its raw trace. It needs the run's results.jsonl and rollout_traces.jsonl in one directory.

Re-run the core evaluation yourself (~$2, one A100-80GB, ~66 min)

pip install vllm==0.27.1
python harness/runpod_g10_eval.py --model LuzCommand/lil-beastie-4b \
    --set decks/_full646_clean.jsonl --out my_run
python harness/score_g10.py --run my_run --boot 300

Use _full646_clean.jsonl (645 questions; the harness draws G=10 per row). _full646_g10.jsonl is the same deck pre-expanded ×10 for the Mac/MLX harness — passing it here would sample each question 100 times.

What's here

path	contents
harness/runpod_g10_eval.py	the G-spread vLLM harness: verbatim prompt with the load-bearing \boxed{IDK} offer, ONE instruction per question chosen from the row (option-letter for multiple choice, free-form otherwise), sampling (temp 1.0 / top_p .95 / top_k 20), last-complete-\boxed{} extraction, letter + numeric-gold + IDK + truncation grading. Traces written unconditionally. (Updated 2026-10-30: the earlier copy appended the option-letter instruction to every row, including free-form decks — see results/aime_ab_brief_q8/FINDING.md.)
harness/score_g10.py	the multiple-choice scorer: per-question vote rollup (trunc/nobox = non-votes, still in the G denominator), consensus curve, ECE (mean-confidence bins) / Brier / rank-AUROC, bootstrap CIs, per-category table.
harness/score_freeform.py	the free-form scorer: outcome rates and commit precision from the graded rows, contamination exclusion, paired t-test between two prompt wordings.
harness/score_humaneval.py	the HumanEval scorer: pass@1 / pass@5, tool-use probes, generation time, paired t-test vs a base run. The HumanEval harness itself runs on the author's MLX trainer stack and is not standalone, so it is not shipped; the per-attempt rows and full traces are.
harness/simpleqa_g10.py	the free-form SimpleQA probe harness.
harness/score_popqa.py, score_qa.py, score_ifeval.py, verify_traces.py	the side-bench scorers: PopQA by popularity band; SelfAware / SQuAD 2.0 / math, with abstention F1; IFEval via lm-eval's official checks; and the trace-to-grade consistency check.
decks/side_benches/	PopQA 1k (10 popularity bands), SelfAware 1k, SQuAD 2.0 1k, math 751 (GSM8K test + MATH-500 integer subset; 0 normalized-text overlap with training), IFEval 541.
results/side_benches/	per-rollout rows and run summaries for the 13 side-bench runs, plus SUMMARY.md and per-bench REPORT_*.md tables.
decks/	every question deck used: the 10,298-question fresh MMLU-Pro deck, the 645 bench deck (_full646_clean.jsonl, and ×10-expanded _full646_g10.jsonl), both AIME decks, SimpleQA (the 1,792 run questions), the q8-check deck.
contamination/	training-id manifests per model (train_ids/*.json, with sha256 of the actual training files), the generated eligibility matrix for the MMLU-Pro decks (0 id-leak, 0 exact-text-leak), the passage-sibling exclusions record (ids 4849, 10565), and aime_training_lineage_ids.json (the 38 AIME problems that appear in the training lineage — see Provenance).
results/<run>/	raw per-rollout outcome rows (results.jsonl) + run summaries for every run on the card: the 33h Mac/MLX 645, the A100 replication, the base-model ablation, both fresh sweeps, the q8 behavior check, SimpleQA (with the semantic grading files), the clean-prompt AIME A/B (aime_ab_brief_q8, aime_ab_through_q8, with instr.txt = the exact instruction, FINDING.md = the prompt artifact, FINDING_CONTAMINATION.md = the lineage overlap), and HumanEval (coding_bench_g5, coding_bench_g5_base with paired.json). Superseded: aime_g10 and aime_full_g5 (A100) — two instructions stacked on a free-form deck halved coverage; kept for the record.
traces/<run>/	full reasoning traces for the runs that have them: Mac-645, fresh2 (all 71,040 rollouts), SimpleQA, q8-check, both AIME A/B arms, both HumanEval runs, all 13 side-bench runs (traces/side_benches/), and the superseded AIME runs. Three early runs are outcome-rows only (an early harness discarded completion text; capture has been unconditional since).

Provenance notes

• The fresh deck = the 11,601-question MMLU-Pro pool minus the 646 bench ids minus every trained id → 10,298; zero id and zero exact-text overlap vs training (see contamination/).
• The card's scoring excludes fresh ids 4849 and 10565 (passage-siblings of trained items, excluded conservatively — documented in contamination/EXCLUSIONS.md).
• AIME: the MMLU-Pro audit above never covered AIME. A later scan found all 32 problems of the curated deck (decks/_aime_g10.jsonl) and 6 of the 64 fresh problems in training pools of ancestor checkpoints — 38 of 96. The card reports AIME on the remaining 58 only; FINDING_CONTAMINATION.md has the counts and the clean numbers; the contaminated slice scores higher than the clean one under the same prompt (the memorization signature).
• Training method and the router/plane system are not part of this release. MMLU-Pro training items are disclosed by id + content hash in contamination/train_ids/; the math training rows (GSM8K, MATH, AIME) are not id-listed there — their overlap with the AIME eval decks is the 38-id list above.