mjf-su/cp-coal-libero90-data

cp_coal LIBERO-90 data bundle

These are the exact inputs of the cp_coal experiment: LoRA fine-tuning of NVIDIA Cosmos-Policy-LIBERO-Predict2-2B on LIBERO-90 with a coalition-averaged action loss. Code, setup and instructions are in https://github.com/mjf-su/wam-influence, branch cp-coal-libero90; start from tr/eval_libero/cp_coal/DATA.md (PATH A).

Files

File	Content
cp_coal_bundle.tar (11.9 GB, uncompressed tar)	Unpacks into <repo>/outputs/. It contains: • cp_coal/cache/train/: 84 shards + index, 42,809 VAE-encoded training examples • cp_coal/{split,moments_heldout,moments_train,moments_summary}.json • cp_future/data/regen/libero_90/: LIBERO-90 demo_0..4 regenerated with NVIDIA's regenerate_libero_dataset.py, 403 successful demos • cp_future/data/t5_libero90.pkl: T5-11B embeddings of the 74 instructions
cp_coal_bundle.tar.sha256	sha256 of the tar
SHA256SUMS	sha256 of every file inside the tar; identical to tr/eval_libero/cp_coal/manifest/bundle_SHA256SUMS in the code branch

Verify and unpack with bash tr/eval_libero/cp_coal/cluster/verify_bundle.sh <download dir>.

Provenance and licences

• LIBERO-90 demonstrations. Derived from yifengzhu-hf/LIBERO-datasets (Apache-2.0) at revision f13aa24, replayed and re-rendered with NVIDIA's Cosmos-Policy regeneration script on stock LIBERO.
• Text embeddings. Computed with google-t5/t5-11b (Apache-2.0).
• Training-cache latents. Computed with NVIDIA's released Cosmos-Policy pipeline and the Wan2.1 VAE tokenizer from nvidia/Cosmos-Predict2-2B-Video2World (NVIDIA Open Model License). Use them in accordance with that licence.