pi05_sortletters_parallel_121_30000_bs2

PI0.5 full fine-tuning checkpoint for RoboTwin sortletters_parallel_121.

Training steps: 30,000
Global batch size: 2
All parameters trainable; AdamW; bfloat16
FSDP devices: 2; seed: 42
Dataset: oss://evo-studio/benchmark/sortletters_parallel_121/
Training code commit: 0361f6652c9a6ac1700fb8c6b24ff166a3beb9f3

This is the complete Orbax checkpoint: params/, train_state/, assets/, and _CHECKPOINT_METADATA. train_state/ contains optimizer state and step, so training can resume with the matching code/config/environment and original two-device FSDP topology. A different device topology requires explicit resharding.

SHA256SUMS is the source artifact manifest; receipts_training_receipt.txt records training and environment details.

Downloads last month: -; Downloads are not tracked for this model. How to track

Video Preview

Robotics