I-BFM: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning
Paper • 2610.06129 • Published
Dataset Viewer
The dataset viewer is not available for this split.
Job manager crashed while running this job (missing heartbeats).
Error code: JobManagerCrashedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
I-BFM 166.560M replay buffer
Replay-buffer snapshot from the I-BFM H200 training run at 166,559,744 environment steps (8 ranks). This dataset accompanies the selected checkpoint in the I-BFM code release and the paper.
Files
checkpoint/buffers/train_rank_0/ through train_rank_7/ each contain buffer.hdf5 and config.json. The eight HDF5 files total 225,689,714,688 bytes. SHA256SUMS records checksums for all 16 buffer files. Preserve both files in each rank directory and the directory names when downloading.
The buffer uses the release code's TrajectoryDictBufferMultiDim HDF5 format. It is a snapshot of replay state, not a standalone training checkpoint; exact resume additionally needs the matching model and optimizer state.
Provenance
- Checkpoint step: 166,559,744 environment steps.
- Training: 8 H200 workers.
- The HDF5 data was generated during the original I-BFM training run. It includes state and action trajectories, latent vectors, and auxiliary reward fields.
- Downloads last month
- 291