AIAAH-1
GRPO fine-tune of openbmb/MiniCPM5-2B for code reasoning.
Training
- Method: GRPO (Group Relative Policy Optimization) via TRL
- Recipe: Inspired by THUDM/Slime
coding_agent_rl(binary test-based reward) - Reward: 1.0 if generated code passes all MBPP unit tests, else 0.0
- Dataset: MBPP (Mostly Basic Python Problems), 150 training problems
- Steps: 150
- Hardware: 1x NVIDIA A100 80GB
Results
| Metric | Value |
|---|---|
| Baseline pass@1 (held-out MBPP) | 30.0% |
| Trained pass@1 (held-out MBPP) | 56.7% |
| Improvement | +26.7 pts (1.89x) |
- Downloads last month
- 31
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support