FastWAM Joint — LIBERO checkpoint (step 21,700)
This repository contains the FastWAM standard Joint training checkpoint at step 21,700. It is a custom PyTorch checkpoint and requires the matching FastWAM code/configuration to load; it is not a standalone Transformers model.
Checkpoint
- File:
checkpoints/fastwam_joint_step_021700.pt - Size: 24,083,156,857 bytes
- SHA-256:
677756a008965a86393f5044e55b4ba6c24b31777f52b298b10544ce35072757 - Original run:
FastWAM/runs/libero_joint_2cam224_1e-4/2026-07-26_12-43-10/
Training configuration
- Base video model: Wan 2.2 TI2V-5B
- Training family: standard FastWAM Joint (not the LingBot-VA initialized run)
- Action expert: 30-layer ActionDiT
- Training resolution/views: 224, two-camera setup
- Learning rate:
1e-4 lambda_action:1- The saved run configuration records resuming from step 4,000.
- The exact training Git branch and commit were not recorded in the run metadata.
Evaluation results
Results depend on the evaluation suite and inference path; they should not be compared without matching those settings.
- Standard LIBERO: Spatial 95.8%, Object 99.4%, Goal 97.0%, LIBERO-10 96.2% (1,942/2,000 pooled episodes).
- Full LIBERO-Plus evaluation: approximately 69.0% with the standard inference path.
- A same-checkpoint one-pass evaluation reported approximately 72.7%; this is an inference-path difference, not a different checkpoint.
- A 15% sampled LIBERO-Plus evaluation reported approximately 72.8%; it is a subset result.
Usage and licensing
The repository currently provides the checkpoint only. Refer to the FastWAM project code and configuration for loading and inference. No license has been added because the applicable release license has not yet been confirmed; repository visibility and usage terms should be set by the owner before publication.