Qwen3-0.6B GSM8K SFT → GRPO
Controlled reproduction of supervised fine-tuning followed by two-GPU GRPO on a fixed GSM8K-100 holdout.
| Stage | Exact numeric accuracy | Required format |
|---|---|---|
| Base | 22% | 0% |
| SFT-160 | 36% | 45% |
| GRPO-60 | 47% | 75% |
GRPO used separate exact-answer and output-format rewards. The default branch
contains the final GRPO model; revision sft-160 contains the SFT precursor,
and tag grpo-60 identifies the default result.
Limits and provenance
This is a 100-question fixed holdout, not a full-benchmark or multi-seed claim. The training corpus is not redistributed here.
- Evidence repository v1.1.0
- Git commit:
0937ef30ea1e3f4b85ec87d1696b22191b62ec14