cosmos3-behavior-r1-checkpoints

Training checkpoints from two in-progress Cosmos3-Nano (NVIDIA cosmos-framework, two-tower Omni-MoT World Foundation Model, ~6.98 B trainable params) joint SFT runs on the BEHAVIOR-1K R1Pro data (23-DOF joint action + world-model video, proprio-state prefix, causal diffusion-forcing, 45056-token packing, 4×B200 FSDP, target 20k iters).

Checkpoints

Subfolder	Run	Iter	Notes
difforce_joint_h65/iter_000010000	JOINT SFT + diffusion forcing (three_way / block_causal)	10000	h65 horizon
difforce_ar_rollout_joint_h65/iter_000006000	JOINT SFT + diffusion forcing + AR-rollout loss	6000	h65 horizon, autoregressive rollout loss

These are the latest finalized checkpoints at upload time (the runs were still training; each latest_checkpoint.txt pointed at these iters).

Format — sharded DCP (not HF / not consolidated)

Each iter_* directory is a PyTorch Distributed Checkpoint (DCP) written by 4-way-FSDP training:

iter_XXXXXXXXX/
  model/      __{0..3}_0.distcp + .metadata   (85 GB — weights)
  optim/      __{0..3}_0.distcp + .metadata   (27 GB — optimizer state)
  scheduler/  __{0..3}_0.distcp + .metadata
  trainer/    __{0..3}_0.distcp + .metadata

This is not a safetensors / HF-loadable checkpoint. It is consumed by the cosmos-framework resume path: place an iter_* dir under the run's IMAGINAIRE_OUTPUT_ROOT checkpoints directory and training auto-resumes from it (exact bit-for-bit resume, since optim/+scheduler/+trainer/ are included). For inference/eval you would first consolidate/convert model/ via the framework's DCP tooling.

Provenance

Built on NVIDIA's cosmos-framework (Cosmos3-Nano). Trained on a re-packaged BEHAVIOR-1K R1Pro LeRobot dataset derived from nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim. Model architecture, weights lineage, and licensing follow NVIDIA's upstream Cosmos + PhysicalAI terms — consult those before use. license: other pending confirmation of redistribution terms.