jeetganatra/qwen3-0.6b-math-star-round1

Qwen3-0.6B · STaR round-1 (math reasoning)

Correctness-gated self-training (STaR / rejection-sampling SFT), round 1, applied to a Qwen3-0.6B math reasoner. Part of the qwen3-nano-math-reasoner post-training comparison.

• Seed: Qwen3-0.6B SFT-distilled from Qwen3-235B-A22B reasoning traces (jeetganatra/qwen3-0.6b-qwen3-235b-a22b-math-distill).
• Method: sample k=4 rollouts/problem over 800 MATH problems (the held-out 500 are excluded), keep only \boxed{}-verified-correct rollouts (1,226 kept from 660/800 problems solved), then SFT on that filtered set.

Results (MATH-500, greedy)

Model	Eval	Accuracy
SFT seed	full-500 @ 4096	37.8% (189/500)
STaR round-1 (this model)	full-500 @ 4096	36.0% (180/500)

Round-1 STaR holds at the seed level (within the ~3 pp two-sample noise floor at n=500), unlike on-policy distillation (OPD/OPSD) which degraded the same seed sharply (to 18% / 16% on a 50-problem subset where the seed scores 50%). At this 0.6B capacity ceiling, correctness-gating prevents regression but does not lift accuracy. Full write-up: ON_POLICY_EXPERIMENTS.md in the GitHub repo.

Weights

Raw PyTorch state dict qwen3-0.6B-star-round1.pth (same format as the seed). Evaluate with the repo's grader:

python src/evaluate_math500.py --runtime hf --which_model reasoning \
  --checkpoint_path qwen3-0.6B-star-round1.pth \
  --dataset_size 500 --max_new_tokens 4096 --eval_batch_size 16