dreamzero-yam-molmoact2-checkpoints

Training-run archive for robocurve/dreamzero-yam-molmoact2 (the merged release; start there). Training code, eval protocol, and model card: robocurve/dreamzero-yam.

Contents:

• checkpoints/step-N/ — all 26 raw LoRA training milestones, weights only (steps 100, 200, 300, then every 500 from 1,000 to 12,000; DeepSpeed optimizer shards are omitted). Do not load these with stock load_lora — it rebuilds the frozen DiT on vanilla Wan2.1 instead of the DreamZero-AgiBot base they were trained against. Use the explicit reconstruction in src/reconstruct.py.
• training-state/checkpoint-12000/ — the final checkpoint with full DeepSpeed optimizer state, resumable for continued training (the val/loss curve had plateaued but not risen at 12k steps). Earlier milestones can be warm-started weights-only via the training repo's --allow-fresh-optimizer path.
• training-state/ top level — trainer state, run config, per-step runtime metrics, and loss logs of the main run.
• renders/ — imagined-vs-real rollout videos from the released checkpoint (model's 2x2 camera-grid dream next to ground-truth frames).