MikhailBird/MolmoAct2-ABC130K-70k

MolmoAct2 ABC130K — step 70,000

Native training checkpoint from full_100k_bs128_20260928_run1, fine-tuned from allenai/MolmoAct2-BimanualYAM on ABC130K with global batch size 128. The checkpoint was saved at step 70,000 of a planned 100,000 updates.

The original checkpoint directory contents are preserved at the repository root:

• config.yaml: saved model and training configuration, including normalization.
• model_and_optim/: 64 distributed model/optimizer shards and .metadata.
• train/: training state for all four ranks.

This is a native MolmoAct2 distributed training checkpoint. It has not been converted to a Transformers from_pretrained export. Use the compatible MolmoAct2 training/checkpoint loader, or convert it before inference as needed.

checkpoint_manifest.json records the byte size and SHA-256 of each original file.