heli-stand/state-rm-sd3.5-medium-pickapicv2

Selected State-RM Checkpoint

This is the selected State-RM checkpoint trained on liuhuohuo2/pick-a-pic-v2, using stabilityai/stable-diffusion-3.5-medium as the backbone.

Configuration

• Checkpoint: checkpoint-final
• T5 maximum sequence length: 256
• Learning rate: 3e-5
• Guidance scale: 5
• Effective batch size: 128
• Resolution: 512
• Sigma range: [0, 1]
• Seed: 42

Final Validation Metrics

Split	Accuracy	Loss	Reward margin
eval	0.6801	0.6152	0.4665
eval_unique	0.6387	0.6126	0.3670

eval was used as the primary selection metric and eval_unique as the complementary metric. This was the best completed State-RM run in the current T5-256 sweep. Its final losses remained stable, with no evidence of overheated training dynamics.

Use this checkpoint for state-aware reward modeling and alignment.