bi199797/vh-features-cosmos-reason2-2b-robometer
vh-features-cosmos-reason2-2b-robometer Frozen-backbone feature cache for the ego-centric progress value head (RLInf-value-head repo), robot side. Naming: vh-features-<backbone>-<data>. Sibling with the Qwen backbone on the same rows: vh-features-qwen3vl-2b-robometer. Backbone: nvidia/Cosmos-Reason2-2B (bf16, frozen). Data: the Robometer processed datasets (robometer/processed_datasets, arXiv 2603.02115): MetaWorld eval+train, LIBERO-10 + LIBERO-10 failure rollouts, MIT Franka… See the full description on the dataset page: https://huggingface.co/datasets/bi199797/vh-features-cosmos-reason2-2b-robometer.
vh-features-cosmos-reason2-2b-robometer
Frozen-backbone feature cache for the ego-centric progress value head (RLInf-value-head repo), robot side. Naming: vh-features-<backbone>-<data>. Sibling with the Qwen backbone on the same rows: vh-features-qwen3vl-2b-robometer.
Backbone: nvidia/Cosmos-Reason2-2B (bf16, frozen). Data: the Robometer processed datasets (robometer/processed_datasets, arXiv 2603.02115): MetaWorld eval+train, LIBERO-10 + LIBERO-10 failure rollouts, MIT Franka, USC Koch / xArm / Franka, UT Dallas SO-101 — 17 sets, 2,540 trajectories × 32 frames, global-task conditioning, target t/T, one wrong-task negative per frame: 162,528 pairs (same pairs.parquet as the Qwen sibling). Quality tiers are in pairs.parquet (quality, success).
Layout (rbm_cosmos/): pairs.parquet, features_index.txt, features_part-NN.tar (uncompressed shards of features/<dataset>__<uuid>.npz, each with pair_id[N] + three [N, 2048] float32 poolings). Frames not included.
Restore: mkdir -p cache/rbm_cosmos/features && for t in rbm_cosmos/features_part-*.tar; do tar xf "$t" -C cache/rbm_cosmos/features; done && cp rbm_cosmos/pairs.parquet cache/rbm_cosmos/