JAM — LIBERO checkpoints

The three original LIBERO paper checkpoints of JAM, one folder per backbone. Each folder is a self-contained package with its checkpoint, normalization statistics, configuration, recorded results, pinned dependencies and release.py.

Project · Code · arXiv: coming soon

Policy	Folder	Checkpoint
JAM (Wan)	libero/wan	12.18 GB
JAM (Ovis-U1)	libero/ovis-u1	4.37 GB
JAM (Qwen3.5-2B)	libero/qwen3.5-2b	7.25 GB

Recorded results

LIBERO: 40 tasks × 50 episodes = 2,000 episodes. LIBERO-Plus: the paper's fixed 1,506-task subset, one episode per task, seed 42; AVG is task-weighted. These are the original single-training-seed results; per-task counts are in each folder's results/.

LIBERO

Policy	Spatial	Object	Goal	Long	AVG
JAM (Wan)	98.80	100.00	98.00	95.80	98.15
JAM (Ovis-U1)	98.60	99.60	98.40	90.40	96.75
JAM (Qwen3.5-2B)	97.80	99.60	97.20	91.60	96.55

LIBERO-Plus

Policy	Spatial	Object	Goal	Long	AVG
JAM (Wan)	74.24	67.20	57.84	59.52	64.54
JAM (Ovis-U1)	82.27	86.24	78.15	73.28	79.95
JAM (Qwen3.5-2B)	84.76	88.10	75.84	76.19	81.14

Download one policy

Resolve the revision once and download one folder of that immutable snapshot:

from huggingface_hub import HfApi, snapshot_download
repo = "y0urOy/JAM"
revision = HfApi().model_info(repo).sha
snapshot_download(repo, revision=revision, allow_patterns="libero/wan/*", local_dir="./jam")
print(revision)  # retain with your experiment

Use libero/ovis-u1/* or libero/qwen3.5-2b/* for the other policies.

Then run python jam/libero/wan/release.py verify and follow that folder's README.md to install the matching environment and evaluate. These policies are loaded by JAM's load_checkpoint, not AutoModel.from_pretrained.

License

Weights retain the Apache-2.0 terms of their base models; see each folder's LICENSE and notice files.