jwjeonn/mastars-dataset

MASTARS offline MARL datasets

Preprocessed offline multi-agent RL datasets used by MASTARS, a data-augmentation method for offline MARL. Everything is stored as NumPy .npy files; drop the contents of this repo into data/ of the MASTARS repo to use them directly.

hf download jwjeonn/mastars-dataset --repo-type dataset --local-dir data

Contents

Env	Scenarios	Qualities	Layout
mpe	simple_spread, simple_tag, simple_world	medium, medium-replay, random	mpe/<scenario>/<quality>/seed_2000_data/{obs,acs,rews,dones}_<agent>.npy
smac	3m, 2s3z, 8m (Good, Medium); 25m, 2c_vs_64zg (Medium)	see left	smac/<map>/<quality>/{obs,actions,rewards,discounts,legals,states,path_lengths}.npy
smacv2	protoss 3_vs_3, 5_vs_5; terran 3_vs_3, 5_vs_5; zerg 3_vs_3	Good, Medium	same as smac

The contents match exactly the datasets used in the MASTARS paper. MPE contains only the seed_2000_data subset (a 2,000-episode subset of each OMAR dataset), matching the small-data setting studied in MASTARS.

Sources

• MPE: OMAR datasets (Pan et al., 2022).
• SMAC: Off-the-Grid MARL (Formanek et al., 2023) tfrecords, converted to concatenated-episode .npy files with scripts/transform_og_marl_dataset.py from the MASTARS repo.
• SMACv2: collected by the MASTARS authors with QMIX (PyMARL), 2,000 episodes per task with a 200-step horizon; Good and Medium follow the paper's quality names.

Please cite the original dataset papers along with MASTARS if you use this data.