Qwen2.5-3B first supervised-fine-tuned on correct multiplication solutions (countdown-mult-sft), then trained with the same GRPO recipe for 300 iterations.

The point of this run was to test whether seeding GRPO with SFT (to install multiplication first) beats GRPO alone. It does not. GRPO restores add/sub that SFT had forgotten (19% back to 75% pass@10), but the multiplication SFT installed is pruned back to 0%, and the rigid SFT template survives, collapsing output diversity to about two distinct answers per ten tries. Stacking them keeps neither half-model's strength.

Full writeup: https://leon2k2k2k.github.io/blog/2026/grpo-sft-teaching-reasoning-through-arithmetic/ Companion: GRPO-alone model | SFT dataset

The model expects the Countdown prompt format: reason inside <think> </think>, give the final equation inside <answer> </answer>.

300 held-out problems (150 add/sub, 150 needs-mult), 10 samples per problem at temperature 0.7.

Compared with GRPO-alone, this model is a touch ahead at a single sample (71% vs 67% add/sub pass@1) but stalls with more tries (75% vs 94% add/sub pass@10): it is committed rather than exploratory.

Two stages, both on one H100. (1) SFT on ~5,000 worked multiplication solutions. (2) GRPO via nano-aha-moment from the SFT checkpoint: G = 4, learning rate 1e-6, KL 0.001, temperature 1.0, 1024-token budget, 300 iterations. Reward = 1.0 well-formed + 1.0 correct.

This is a fine-tune of Qwen2.5-3B by the Qwen team, and is released under the same Qwen Research License. The base model and its weights are their work; this repo only adds SFT then GRPO fine-tuning on Countdown.