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.
