HumorR1/policy-qwen3vl-2b-grpo-newyorker

Add 8192-token eval JSON

Update model card with honest 8192-token eval (was clipped at 2048)

Add eval JSON for the SFT+GRPO policy

Replace with SFT-think + GRPO policy (+2.30σ over base, restores thinking format)

Add 50×5 held-out eval (RM scores per cartoon × sample)

GRPO policy (Qwen3-VL-2B-Thinking + LoRA, FB=0.1, 200 steps)

initial commit