---
license: apache-2.0
base_model: hemlang/Hemlock2-Coder-7B
datasets:
  - hemlang/hemlock-codex3-SFT
tags:
  - hemlock
  - code
  - grpo
  - reinforcement-learning
---

# Hemlock2-Coder-7B-GRPO

[hemlang/Hemlock2-Coder-7B](https://huggingface.co/hemlang/Hemlock2-Coder-7B) improved with
execution-reward GRPO ([grimoire](https://github.com/Schneewolf-Labs/grimoire) ≥ 2.0 /
[hemlock-rl](https://github.com/Schneewolf-Labs/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](https://github.com/hemlang/hemlock/tree/main/benchmark)
(zero-shot, n=5, benchmark-overlapping training tasks held out):

| | pass@1 | pass@5 |
|---|---|---|
| Hemlock2-Coder-7B | 28.9% | 55.3% |
| **Hemlock2-Coder-7B-GRPO** | **36.8%** (+7.9) | **57.9%** (+2.6) |

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](https://huggingface.co/datasets/hemlang/hemlock-codex3-SFT)),
  one language-variant per task, hembench-name-overlapping tasks excluded.

Q8_0 GGUF included.