AIDA PI0 CALVIN ABC SFT — step 13,500
This is an interim, full-parameter PI0 supervised-finetuning checkpoint trained on the ABC portion of the CALVIN dataset in LeRobot format. It was exported from the ongoing 30,000-step AIDA/RLinf training run at step 13,500.
Checkpoint summary
- Base model:
lerobot/pi0_base(25c379b52ba2ff8788cab921758a3cc3fe3f77f2) - Dataset:
InternRobotics/InternData-Calvin_ABC(fd844a43912c1823ff8355873b6f013a5cce127a) - Training: full-parameter SFT, global batch size 16, micro-batch size 1
- Progress: 13,500 of 30,000 planned optimizer steps
- Fixed ABC validation loss at this checkpoint: 0.106
- Validation set: one deterministic midpoint frame from each of 799 held-out ABC episodes
- Action contract used by the training runtime: five future 7D relative actions
- Normalization: statistics computed only from the 15,177-episode training
split, bundled under
InternRobotics/InternData-Calvin_ABC/norm_stats.json
CALVIN D was not mounted, inspected, or used for training or checkpoint selection.
Files
model.safetensors: standalone PI0 model weightsconfig.json: LeRobot PI0 policy configurationpolicy_preprocessor.jsonandpolicy_postprocessor.json: policy processor configurationInternRobotics/InternData-Calvin_ABC/norm_stats.json: train-split normalization statisticsexport_manifest.json: export hashes and state-dictionary compatibility detailscheckpoint_metadata.json: sanitized training and provenance metadata
Provenance
- RLinf source revision:
0f9ea98c7a6d9e3ade24e8f4846c64d3b135dbcc - Training runtime image:
sha256:d9eaa92d93aeba209b2af1ec810cb8bf866d432d27f160c7ab803d3e6274aadb - Exported model SHA-256:
1ec9c2c3abb0c92b735bd125d297257f697737de623677e3e9de552de330143b - Source RLinf full-weights SHA-256:
c03a582a86bca5896909392206674feff0c615e7815aefaeb98f6e9337f302fb - Train-split normalization SHA-256:
abd6c2f8228d04e1677e86fee497e15477b1d4e5726e334373a045ad161af29a
Limitations
This is not the final checkpoint. The full training run is still in progress, and this checkpoint has not yet been selected using final-run comparisons or evaluated on the CALVIN task-success benchmark. The reported validation value is the fixed held-out ABC flow-matching loss, not a task success rate.
Use is subject to the base model's Gemma license and the terms of the referenced dataset and training software.