inspect_evals/gsm8k, full 1319-sample test set, 10-shot, greedy: 0.7081 +/- 0.0125 (untrained Qwen/Qwen3-0.6B baseline: 0.4754 +/- 0.0138).

A statistical tie with the campaign's best model, cmpatino/qwen-grpo-r5 (0.7089 +/- 0.0125) -- see that model card for the full recipe, reward function, and caveats. Note this is the step-125 checkpoint, which scores higher than run r4's final step-129 weights (0.6907).

Runs in non-thinking mode: the chat template always emits an empty <think></think> block. Use the tokenizer shipped in this repo.