NeurIPsMay1234/openpi-interpretability-data

openpi-interpretability-data数据集包含从开放视觉语言动作（VLA）策略模型中提取的可解释性工件，这些模型在LIBERO、MetaWorld和RoboCasa基准上进行评估。工件包括每步层激活数据（从rollouts中提取）、概念器矩阵、线性转向向量、稀疏自编码器向量及相关检查点。数据集支持匿名提交，并用于双盲同行评审。具体涵盖的模型包括π系列VLA（如pi0_5和pi0_fast）和NVIDIA GR00T（如GR00T-N1.5），每个模型对应不同的基准任务。数据以压缩文件格式（如.tar和.npz）组织，需解压后使用，并提供了Python加载示例。许可证为MIT。openpi-interpretability-data is a dataset containing interpretability artifacts (activations, conceptors, linear steering vectors, sparse autoencoder vectors, and checkpoints) extracted from open vision-language-action (VLA) policy models on the LIBERO, MetaWorld, and RoboCasa benchmarks. It accompanies an anonymous submission and is shared for double-blind peer review. The dataset includes models such as the π-series VLA (e.g., pi0_5 and pi0_fast) and NVIDIA GR00T (e.g., GR00T-N1.5), each associated with specific benchmarks. Data is organized in compressed formats (e.g., .tar and .npz files) that require extraction before use, with Python loading examples provided. The license is MIT.