Qwen3-8B GRPO GSM8K
Qwen3-8B trained with GRPO and LoRA using veRL v0.9.1. LoRA weights have been merged into the model.
Training
- Hardware: 2 × NVIDIA A800 80GB
- Training steps: 116
- LoRA rank: 64
- LoRA alpha: 32
- Selected checkpoint: global_step_116
Evaluation
Metric: val-core/openai/gsm8k/acc/mean@1
| Checkpoint | Accuracy |
|---|---|
| Initial model | 24.94% |
| Step 100 | 68.99% |
| Step 116 | 72.33% |
Step 116 achieved the highest validation accuracy among retained checkpoints. Step 115 reached 72.71%, but its weights were not saved.
These scores were recorded before export under the training recipe's evaluation settings. The GSM8K test split was used as validation for checkpoint selection.
See training_summary.json for provenance.