# ChainFlow-VLA VLM Feature Cache This dataset provides cached VLM hidden features and trajectory targets used by ChainFlow-VLA on NAVSIM. ## Dataset Structure navtest/ navtest.tar.zst checksums_navtest.sha256 trainval/ trainval_0000.tar.zst trainval_0001.tar.zst ... trainval_0023.tar.zst checksums_trainval.sha256 trainval_logs.txt The `trainval` split is sharded by NAVSIM log folders. Each shard is an independent `.tar.zst` archive. After extraction, each token folder contains: internvl_feature.gz trajectory_target.gz ## Extract Extract navtest: tar --zstd -xf navtest/navtest.tar.zst Extract all trainval shards: for f in trainval/trainval_*.tar.zst; do tar --zstd -xf "$f" done ## Verify Checksums Verify navtest: cd navtest sha256sum -c checksums_navtest.sha256 Verify trainval: cd trainval sha256sum -c checksums_trainval.sha256 ## Notes - `trainval` contains 24 independent shards. - `navtest` is packed into a single archive. - The archives use `zstd` compression. Please install `zstd` before extraction if your system does not support `tar --zstd`.