dddddsds31341/Long-WAM-Sweep-Trash-Epoch-3

Long-WAM Sweep Trash — Epoch 3

Evaluation weights from epoch 3 of 5, optimizer step 3033, of the independent sweep-trash fine-tune. The run starts from Efficient-Large-Model/Long-WAM-YAM; it does not continue from another task's fine-tune.

Task instruction

Use one gripper to grasp and hold the dustpan, while the other grasps the brush and sweeps both pieces of crumpled paper into the dustpan.

Training setup

• Dataset: yuyingyinya/sweep_trash, revision 0dd5ff1f576de17dcf06b49d352e0f125a1f43e6.
• 100 demonstrations, 48,495 frames, 30 Hz; all demonstrations are used for training.
• Four GPUs, batch size 12 per GPU, gradient accumulation 1, global batch 48.
• Constant learning rate 1e-5, AdamW, weight decay 0.01; BF16 and DeepSpeed ZeRO-1.
• 1,011 optimizer updates per epoch; the full run targets 5,055 updates.
• 14-dimensional absolute joint targets with original gripper units, ordered left 6 joints + gripper, right 6 joints + gripper.
• Task-specific z-score statistics and the official [-5, 5] normalized-value clipping.
• Three RGB cameras: top_head, hand_left, hand_right; processed video layout 384×320, P48 history, 32-step action chunk.

Included files

• model.pt: joint video/action model and proprio encoder, stored step and dtype. No optimizer or scheduler state.
• config.yaml: the resolved configuration captured by the training run.
• dataset_stats.json: this task's normalization statistics; use these for preprocessing and action denormalization.
• code.zip: code-only SIP release v22 used for the training run, based on NVlabs/LongLive@cfa4e91eec59429d583326500bf5535c9ce6eec7 with the SIP changes included.
• requirements-sip.txt: recorded package versions for the SIP environment.

Loading and evaluation

Extract code.zip and install the included Long-WAM checkout. Use its longwam.inference.load_run_config and load_model with the explicit model.pt path, expected_step=3033 and strict=True. Model configuration and normalization statistics must come from this same repository.

The original training configuration retains SIP filesystem paths. For a different machine, set the model-assets root via LONGWAM_MODEL_ROOT and update data/statistics/cache paths for the intended evaluation entry point. Wan2.2 VAE, UMT5 text encoder and tokenizer assets are required separately. This repository is not a standalone Transformers from_pretrained model.

For deployment, use a matching robot adapter and the original 14D action/gripper conventions. This checkpoint has passed CPU file-loading checks; no closed-loop robot success rate or held-out validation result is reported. Training loss is not a robot-task success metric. The eval-only file cannot restore optimizer state for exact training continuation.

Online training record: https://forge.coreweave.com/wandb/bofang/long-wam/runs/nwa2ucgt

Base model and its published license metadata: https://huggingface.co/Efficient-Large-Model/Long-WAM-YAM