finetuned_with_lift_pi05
Private backup of a Pi0.5 Lift behavior-cloning fine-tune checkpoint.
Base
- Base model path during training:
/data/ubuntu/models/pi05/pi05_base
- Policy type: LeRobot PI05Policy / Pi0.5-style model
Dataset
- Source: MIML-VLA Lift HDF5
- Converted dataset path:
/data/ubuntu/miml_vla/suturing_pi05/data/pi05_lift_32d
- Total samples: 2368
- Train samples: 2132
- Validation samples: 236
- Original state: 24D
- Padded state: 32D
- Original action: 7D
- Padded action: 32D
- Action chunk size: 50
Training
- Steps: 100
- Trainable mode: last_layers
- Trainable parameters: about 263M / 4.14B
- Batch size: 1
- Learning rate: 1e-6
- Final train loss: 0.025313330814242363
- Internal validation loss: 0.057169350981712344
Notes
This is the behavior-cloning fine-tuned checkpoint before online RLT. It is intended as the starting checkpoint for rlt_finetuned_pi05.
