logan7000/grpo-qwen3-1p7b-math345-end

grpo-qwen3-1p7b-math345 · end checkpoint (step 136)

Fine-tuned from Qwen/Qwen3-1.7B-Base with Vanilla GRPO against the dataset's ground-truth solutions.

• Training data: q1716523669/MATH-Level345 (8,740 problems, MATH levels 3-5)
• Eval: MATH-500 (500 prompts, pass@1, temperature 0.6)
• This checkpoint: step 136 of 136 (2 epochs) — the end one
• Its score: 0.6706

best is the step with the highest eval; end is step 136. Both are published because they answer different questions and the gap between them is a result in itself. The sibling repo is grpo-qwen3-1p7b-math345-best.

Hyperparameters

Effective batch 128 prompts per optimizer step, lr 3e-6, 2 epochs, 12 generations per prompt, 3072-token completions, temperature 1.0 for rollouts and 0.6 for eval, beta 0, bnpo loss, scale_rewards group, seed 42.

Eval curve

step	MATH-500 pass@1
0	0.5437
10	0.6131
20	0.6290
30	0.6508
40	0.6706
50	0.6746
60	0.6766
70	0.6488
80	0.6845
90	0.6845
100	0.6607
110	0.6647
120	0.6667
130	0.6706

Files for audit

• train.log — the complete training log this checkpoint came from
• eval_curve.csv — the table above, machine-readable
• run_config.json — resolved config as the trainer saw it

Code: williamium3000/trl-projects, branch n3-interaction-modes, projects/co-grpo-dp/.