evo × LawBench — gpt-oss-120b LoRA (Chinese criminal charge prediction)
LoRA adapter (r=32, alpha=64) on openai/gpt-oss-120b, trained on the LawBench
191-class charge-prediction train split (5,332 examples) on 4× H100, produced by an
autonomous evo run (evo 0.5.0-alpha.13; whole-system optimize loop with the
concurrent meta controller).
Result (LawBench 191-class, 913-case held-out test, top-1 accuracy)
| approach | acc |
|---|
| prior SOTA | 0.450 |
| SIA gpt-oss-120b (W+H) | 0.701 |
| TF-IDF harness (no LLM) | 0.760 |
| best: this LoRA (vLLM, rope_theta fix) ⊕ TF-IDF ensemble | 0.77 |
Beats SOTA and SIA's 0.701. Honest caveat: the winning 0.77 is an ensemble
of this LoRA (served via vLLM with a rope_theta fix) and a TF-IDF char-ngram
classifier; the LoRA's contribution is the marginal lift over the 0.760 harness.
Naive LoRA inference without the rope_theta fix scored far lower.
Files
adapter_model.safetensors — the LoRA weightsadapter_config.json, tokenizer.json, chat_template.jinja
Replicate
Task + harness: github.com/evo-hq/evo-posttrainbench (branch evo-variant,
src/eval/tasks/lawbench). Optimizer: evo 0.5.0-alpha.13 (github.com/evo-hq/evo,
release/0.5). Run: scripts/run.sh run lawbench openai/gpt-oss-120b <hours>.
