model card: 100k final

README.md

@@ -5,11 +5,15 @@ pipeline_tag: robotics
 tags: [robotics, world-model, video-prediction, fastwam, wan2.2, robodojo]
 ---
 
-# FastWAM — RoboDojo Precision suite, vertical 576x256 (step 80k)
+# FastWAM — RoboDojo Precision suite, vertical 576x256
 
 FastWAM (Wan2.2-TI2V-5B backbone) trained on the RoboDojo **Precision** suite, 8 tasks.
-Intermediate checkpoint at **step 80,000 of 100,000** — the run was still going when this
-was exported.
+The full 100,000-step run, plus the 80k checkpoint kept for comparison.
+
+**Use `step_100000.pt`.** The cosine schedule anneals hard at the end: val_loss went
+0.1011 at 80k to **0.0272** at 100k. The 80k file was uploaded while the run was still in
+flight and is kept only so the difference is checkable, not because it is a useful
+alternative.
 
 ## Training setup
 
@@ -17,12 +21,12 @@ was exported.
 |---|---|
 | base model | `Wan-AI/Wan2.2-TI2V-5B` |
 | ActionDiT backbone | linear-interpolated from the Wan2.2 DiT, alpha-scaled, 1024 hidden |
-| steps | 80,000 (of 100,000) |
+| steps | 100,000 |
 | GPUs | 8x H100 80GB, DeepSpeed ZeRO-1 |
 | batch | 8 per GPU = **global batch 64** |
 | lr | 1e-4 cosine, weight decay 1e-2 |
 | grad accumulation | 1 |
-| loss at export | total 0.0803, action 0.0006 |
+| val_loss | **0.0272** at 100k (0.1011 at 80k) |
 
 ## Canvas and cameras
 
@@ -57,7 +61,8 @@ insert_key, insert_tubes, play_Xylophone, plug_in_charger, pour_balls_into_vase.
 ## What is in here
 
 ```
-step_080000.pt       weights only (12 GB) -- not the ZeRO training state
+step_100000.pt       final weights (12 GB) -- use this one
+step_080000.pt       intermediate, kept for comparison
 dataset_stats.json   z-score normalization computed over the 800-episode subset
 config.yaml          the resolved training config for this run
 ```
@@ -69,7 +74,9 @@ included — this checkpoint is for inference, not for resuming.
 
 ## Notes
 
-- This is **step 80k of a 100k cosine schedule**, so the LR had not annealed to its floor.
-  A later checkpoint should be better; compare rather than assume.
 - Trained with `mot_checkpoint_mixed_attn: false` and the uncond (no action conditioning
   on the video DiT) recipe.
+- The run was interrupted once at step ~58.9k and resumed from the 55k **state**
+  checkpoint — optimizer moments, LR schedule position and step counter all restored, so
+  the cosine schedule is intact. Resuming from weights alone would have restarted the
+  schedule and is not what happened here.