nbeerbower/Hemlock2-Coder-7B-GRPO

Hemlock2-Coder-7B-GRPO

hemlang/Hemlock2-Coder-7B improved with execution-reward GRPO (grimoire ≥ 2.0 / hemlock-rl). Completions were executed in the Hemlock interpreter's sandbox and rewarded for exactly reproducing verified reference stdout.

Strictly improves the base model on hembench (zero-shot, n=5, benchmark-overlapping training tasks held out):

Largest gains in syntax (L1 pass@1 2/9 → 4/9, pass@5 7/9 → 8/9) and systems/concurrency (L4 pass@1 1/7 → 3/7).

Training

• —Method: GRPO (LoRA r=16 α=32, β=0.1 KL vs frozen base, G=8, dynamic sampling, unscaled advantages), 186 deduplicated prompts × 3 epochs.
• —Rewards: sandboxed execution; exact stdout match vs verified reference = 2.0, runs-but-wrong = 0.5, error = −0.5, no code / timeout = −1.0.
• —Data: execution-verified translation/generation tasks (now published as hemlang/hemlock-codex3-SFT), one language-variant per task, hembench-name-overlapping tasks excluded.

Q8_0 GGUF included.