frmjua/moltenpot

Molten Pot — Offline Multi-Agent RL Datasets

Logged focal-agent trajectories for the Molten Pot benchmark: offline multi-agent reinforcement learning in mixed-motive (social-dilemma) settings, built on five DeepMind Melting Pot substrates. This dataset accompanies the paper. See the paper and the code repository for full methodological detail; this card documents the released data itself.

Dataset version: v1.0 — released

License: the datasets are released under CC-BY-4.0. You may share and adapt the data for any purpose, including commercially, provided you give appropriate credit (cite the accompanying paper) and indicate if changes were made. Full text: https://creativecommons.org/licenses/by/4.0/legalcode. The benchmark's code is released separately under the MIT License. The environments themselves are derived from DeepMind Melting Pot (Apache-2.0).

---

1. Directory layout and file naming

Files are organised nested by substrate, then by scenario:

<substrate>/<scenario>/dataset.hdf5

For example, clean_up/clean_up_0/dataset.hdf5 holds the logged data for scenario 0 of the Clean Up substrate. Scenario names match the paper and the benchmark code exactly (Section 2 below lists every scenario name).

Each dataset.hdf5 also has a stats.json sibling file with summary statistics for that scenario's episode returns (mean, std, min, max, median, p25/p75/p90), used to compute the D4RL-style normalisation in the paper.

2. File format and field definitions

Each dataset.hdf5 is an HDF5 file with one top-level group named after the scenario, containing one subgroup per logged episode (ep_0, ep_1, ...).

Field	Shape	dtype	Description
obs	(T, A, 3, 88, 88)	uint8	Egocentric RGB observation per focal agent per timestep, channels-first, 0–255 (LZF-compressed on disk; normalise to [0, 1] at load time).
actions	(T, A)	int32	Discrete action index taken by each focal agent at each timestep (see the paper's substrate table for each substrate's action-space size).
rewards	(T, A)	float32	Per-agent scalar reward at each timestep.
dones	(T,)	bool	Episode-termination flag, shared across agents.

T is the episode length (1,000 timesteps for every scenario in this release), and A is the number of focal agents in that scenario (ranges from 1 to 14 across substrates — see Section 2).

Collection process. Trajectories were generated by training independent PPO behaviour policies per scenario and logging episodes uniformly across the entire course of training, from random initialisation through convergence. Each scenario's dataset is therefore a skill-mixed cross-section rather than expert-only demonstrations.

Loading example:

import h5py
import numpy as np

with h5py.File("clean_up/clean_up_0/dataset.hdf5", "r") as f:
    scenario = f["clean_up_0"]
    ep = scenario["ep_0"]
    obs     = ep["obs"][:]      # (T, A, 3, 88, 88) uint8
    actions = ep["actions"][:]  # (T, A) int32
    rewards = ep["rewards"][:]  # (T, A) float32
    dones   = ep["dones"][:]    # (T,) bool
    h_states = ep["h_states"][:]  # (T+1, A, H) float32

obs = obs.astype(np.float32) / 255.0

3. Scenario catalogue (47 scenarios, 5 substrates)

Substrate	# Scenarios	Scenario names
Clean Up	9	clean_up_{0,2,3,4,5,7,13,14,15}
Coins	7	coins_{0,1,2,3,4,5,6}
Coop Mining	6	coop_mining_{0,1,2,3,4,5}
Commons Harvest	12	commons_harvest__closed_{0,1,2,3}, commons_harvest__open_{0,1}, commons_harvest__partnership_{0,1,2,3,4,5}
Allelopathic Harvest	13	allelopathic_harvest__open_{0,1,2,3,4,5,6,7,8,9,10,11,12}

Full per-scenario descriptions (social role, focal-population type) are in the paper's benchmark appendix and on the project website's split visualiser.

4. Train/test split mapping (Evaluation Setting 3)

Evaluation Setting 3 evaluates zero-shot social generalisation: a policy is trained only on a split's train scenarios and evaluated, without further adaptation, on its held-out test scenarios. Every split below uses only files already listed in Section 3 — no additional data. The files used for a given split's training run are exactly its "Train scenarios" column; the files used for its zero-shot evaluation are exactly its "Test scenarios" column.

Coins

Split	Name	Train scenarios	Test scenarios
C1	The Kindness Trap	coins_{0,5,6}	coins_{1,2,3,4}
C2	Hair Trigger	coins_{1,3}	coins_{2,4}
C3	Escalation Protocol	coins_{1,2}	coins_{3,4}
C4	Wild Card	coins_{5,6}	coins_{0}

Clean Up

Split	Name	Train scenarios	Test scenarios
U1	Terms and Conditions	clean_up_{0,2,3,7}	clean_up_{13,14,15}
U2	Quid Pro Quo	clean_up_{0,2,3,15}	clean_up_{4,5,7}
U3	Home Turf	clean_up_{0,2,3,7,13}	clean_up_{4,5}

Coop Mining

Split	Name	Train scenarios	Test scenarios
M1	Home and Away	coop_mining_{0,1,2,5}	coop_mining_{3,4}
M2	Signal in the Noise	coop_mining_{0,2}	coop_mining_{5}
M3	Scarcity of Trust	coop_mining_{0,5}	coop_mining_{3}

Commons Harvest

Split	Name	Train scenarios	Test scenarios
H1	Paper Tigers	commons_harvest__closed_{0,1}, commons_harvest__open_1, commons_harvest__partnership_4	commons_harvest__open_0, commons_harvest__partnership_5
H2	Majority Rules	commons_harvest__closed_{0,2}, commons_harvest__partnership_{0,2,4,5}	commons_harvest__closed_{1,3}, commons_harvest__open_{0,1}, commons_harvest__partnership_{1,3}
H3	The Social Contract	commons_harvest__closed_{0,1,2,3}, commons_harvest__open_{0,1}	commons_harvest__partnership_{0,1,2,3}
H4	Good Faith, Bad Outcome	commons_harvest__partnership_{0,1}	commons_harvest__partnership_{2,3}

Allelopathic Harvest

Split	Name	Train scenarios	Test scenarios
A1	Against the Grain	allelopathic_harvest__open_{1,3,4,7,8}	allelopathic_harvest__open_{0,5,6}
A2	The Bandwagon Effect	allelopathic_harvest__open_{3,7,8,9,10,11}	allelopathic_harvest__open_{4,5,6,12}
A3	Small World	allelopathic_harvest__open_{3,4,5,11,12}	allelopathic_harvest__open_{6,7,8,9,10}
A4	From Follower to Leader	allelopathic_harvest__open_{0,1}	allelopathic_harvest__open_{2,9,10,11,12}

Note on Evaluation Settings 1 and 2. These two settings do not use train/test splits. Setting 1 trains one independent policy per scenario (evaluated on that same scenario); Setting 2 pools every scenario in a substrate into a single training set (the scenario label withheld at training time) and evaluates on each scenario in that same pooled set.

5. Uses and limitations

Intended for training and evaluating offline-RL algorithms under the three evaluation settings above, and for studying social outcomes (per-agent return inequality, background-agent welfare, generalisation gap) in mixed-motive multi-agent settings. The datasets describe behaviour in abstract grid-world social dilemmas and should not be treated as models of real human social behaviour or used to make claims about specific human populations. All trajectories are generated entirely in simulation; there is no human, personal, or otherwise sensitive data.

6. Citation

If you use this dataset, please cite:

@inproceedings{TODO}

7. Maintenance

Maintained by the authors of the accompanying paper. Corrections and additional scenarios may be released as new dataset versions on the Hub; this card's Dataset version field and the Hub's commit history are the source of truth for what a given release contains. Please file issues or questions on the project's GitHub repository.