01 — Work / 05LLMs · ML← All projects

RL fine-tuning for code generation

GRPO fine-tuning of Qwen2.5-Coder with unit-test rewards and a sandboxed executor. Lifted 1.5B pass@1 by +3.2 on MBPP and +5.5 on HumanEval+ (3 seeds), with no reward hacking across 21 audited runs.

Type: Build
Tracks: LLMs · ML
Context: TRL · LoRA · vLLM
Links: Code ↗

greedy pass@1 of Qwen2.5-Coder-1.5B before and after GRPO, mean of 3 seeds

FIG. 05