GRPO where the model's own 0-9 rating is its reward. 20 runs (2 follow-ups x seeds 21-30); every checkpoint is a branch, step50 to step600 (= main).
-
fjzzq2002/qwen3.5-2b-self-reward-grading-s21
Text Generation • 2B • Updated • 70 -
fjzzq2002/qwen3.5-2b-self-reward-grading-s22
Text Generation • 2B • Updated • 77 -
fjzzq2002/qwen3.5-2b-self-reward-grading-s23
Text Generation • 2B • Updated • 73 -
fjzzq2002/qwen3.5-2b-self-reward-grading-s24
Text Generation • 2B • Updated • 77
