pi0.5 TACO LIBERO-32 Checkpoints
This repository archives the 36 checkpoints referenced by the final Pretrain-32 LIBERO evaluation manifest. It includes the pretrained model and the SFT, RETAIN (alpha=0.9), LoRA, legacy EWC, L2 All, TACO All, and standard WTACO checkpoints for the single-task and sequential experiments.
Each checkpoint is stored under checkpoints/<tag>/. Full training state is
retained when it exists so that experiments can be resumed. manifest.tsv
maps each tag to its experiment, method, OpenPI config, and focus task IDs.
The associated 5-seed x 10-episode evaluation artifacts are stored in the
Berkeley-ICON-Lab/pi05-taco-libero32-evals dataset repository.
These checkpoints use the Orbax layout expected by the accompanying OpenPI code. See the evaluation repository for the exact scripts and code revision.