ws-wm-0218-step-120

Exp8-r2 (0218) World Model checkpoint at step 120 (epoch 0.86).

Training

• Method: Pivot-GRPO with Cauchy α=1.0, BehR-only reward
• Base: Qwen2.5-7B SFT World Model
• Bug Fixes: 9 critical fixes (negative reward, signal dilution, API dedup, etc.)
• Ckpt Monitor BehR: See collection description for full results

Key Config

• reward_mode=cauchy, behavior_scale_coef=1.0
• behavior_weight=1.0, facts_weight=0, length_penalty_weight=0
• lr=2e-6, batch_size=32, temperature=1.2
• max_prompt_length=28672, ppo_epochs=1
• Judge: Qwen3-8B (vLLM TP=4)

Part of collection: ws-wm-0218