# eval3_wallx

Wall-X fine-tune for Eval 3: place an empty 330 ml slim coke can on the celebrity image named in the prompt.

Important: this upload is the available `003000` checkpoint from the local run. A `005000` checkpoint was not present on disk.

Training summary:

- Base model: `x-square-robot/wall-oss-flow`
- Policy: LeRobot `wall_x`
- Dataset requested: `robot-learning-group47/eval3_180_new`
- Local training subset: first 99 complete episodes from `eval3_180_new`
- Reason for subset: the Hub dataset has 180 parquet episodes but only 104 MP4 files; LeRobot requires videos for selected episodes.
- Steps: 3000
- Batch size: 4
- Effective batch size: 4
- Device: A100 80 GB
- W&B: offline