This repository contains 94,372 multimodal robot subgoal training samples from 16 tasks and the 240,855 unique PNG files referenced by those samples. Images are distributed as 16 lossless tar.zst archives, one per task. Extract the archives before training; the image paths in the JSONL point to the extracted files. Install the Hugging Face CLI (pip install -U huggingfacehub) and the zstd command (brew install zstd on macOS or sudo apt-get install zstd on Debian/Ubuntu). Then, from an empty directory: The helper requires Python 3.9 or later. It verifies archive SHA-256 checksums, validates every member path, and verifies all restored PNG files against the image manifest. It resumes safely…