JackLiu0406/b1k-checkpoints

B1K Checkpoints

This repository contains BEHAVIOR-1K pi0.5 checkpoints produced from the first-place 2025 B1K model code path.

task31_clean_boxing_gloves_pi05_baseline_20k_step19999_params_only

Evaluation checkpoint for task 31, clean_boxing_gloves.

• Architecture: BEHAVIOR-1K first-place pi0.5 / pi_behavior model.
• Initialization: IliaLarchenko/behavior_50t_checkpoint.
• Fine-tuning data: task 31 only, all 200 available episodes in the local task dataset.
• Training target: flow-matching action velocity prediction.
• Requested training length: 20k iterations.
• Saved step: 19999 because the training loop saves zero-indexed steps.
• DA3/spatial add-on: not included and not required for loading this checkpoint.
• Contents: params/ plus assets/ only. This is intended for rollout/evaluation, not exact optimizer-state resume.

Use the same evaluator/loader you already use for IliaLarchenko/behavior_50t_checkpoint, but point the checkpoint root at:

task31_clean_boxing_gloves_pi05_baseline_20k_step19999_params_only

The folder layout mirrors the original params-only checkpoint style:

task31_clean_boxing_gloves_pi05_baseline_20k_step19999_params_only/
  assets/
  params/

Do not look for train_state/; it was intentionally omitted to keep the evaluation artifact small.