vikpant/coopetition-gym-logs

Coopetition-Gym v1, Training Results

Training results from the Coopetition-Gym v1 benchmark campaign. 17,930 JSON files, each recording the outcome of training one of 16 reinforcement learning algorithms, 7 game-theoretic oracles, 2 heuristic baselines, or 101 constant-action policies on one of 20 mixed-motive multi-agent environments under one of three reward configurations (private, integrated, cooperative) with one of seven random seeds.

Companion technical report: Coopetition-Gym v1: A Formally Grounded Platform for Mixed-Motive Multi-Agent Reinforcement Learning under Strategic Coopetition. Pant and Yu, arXiv:2605.02063, 2026.

Companion code: https://github.com/vikpant/strategic-coopetition

Companion audit dataset: https://huggingface.co/datasets/vikpant/coopetition-gym-logs

---

Quick Start

pip install huggingface_hub
huggingface-cli download vikpant/coopetition-gym-logs \
    --repo-type dataset --local-dir data/ \
    --include "training_runs/*"

The training corpus is delivered as 950 JSONL shards (training_runs_NNNN.jsonl, ~5 MB each) under data/training_runs/. Each line of each shard is one training-run record (one JSON object per training experiment). Shards can be read line-by-line without an extraction step:

import json
from pathlib import Path

for shard in sorted(Path("data/training_runs").glob("*.jsonl")):
    with open(shard) as fh:
        for line in fh:
            record = json.loads(line)
            # record contains: algorithm, environment, training_seed,
            # status, training_time_seconds, evaluation_time_seconds,
            # metrics, timestamp, gpu_id, tr_mode

Schema

Each JSON file has the following top-level structure:

Field	Type	Description
algorithm	str	Algorithm name (e.g., ISAC, COMA)
environment	str	Environment ID (e.g., TrustDilemma-v0)
training_seed	int	Seed in {99, 100, 101, 102, 103, 104, 105}
status	str	success for released files
training_time_seconds	float	Wall-clock training time
evaluation_time_seconds	float	Wall-clock evaluation time
metrics	object	Aggregated per-episode and per-seed metrics
timestamp	str	ISO 8601 completion timestamp
gpu_id	int	GPU index used (−1 for CPU)
tr_mode	str	TR tier label (tr1, tr2, tr3, tr4)

The nested metrics object contains mean_return, std_return, mean_cooperation_rate, mean_final_trust, training_returns, training_timesteps, training_metrics (gradient-level diagnostics by step), and TR-tier-specific tr_metrics.

See the Croissant metadata for the complete machine-readable schema with JSONPath extractions.

Known Data Characteristics

• Retry-cycle deduplication. The training corpus shipped here is the deduplicated form of the underlying experimental campaign. The campaign produced 27,613 raw run records; 9,683 of these were retry-cycle duplicates (the same (algorithm, environment, seed, tr_mode) cell completed multiple times because the experiment runner retried on transient cloud-instance failures). The deduplication policy is keep latest-timestamp record per cell, which retains the run that completed successfully on its final attempt; the resulting corpus is 17,930 unique cells. All paper analyses, tables, figures, and headline counts are computed on this 17,930-record deduplicated dataset. The deduplication step is a single logical operation applied uniformly across all 950 JSONL shards, which is why every shard's most-recent commit message reads Deduplicate retry runs: keep latest-timestamp per (algo, env, seed, tr_mode); 27613 -> 17930 records.
• 62 NaN-return files from documented training instabilities are retained for transparency: 21 MASAC on TR-3 environments under baseline, 21 MADDPG /MATD3/M3DDPG on ApacheProject-v0 under cooperative reward, 20 MADDPG on AppleAppStore-v0 in the network sensitivity analysis.
• MeanFieldAC is evaluated only on environments with ≥ 3 agents (12 of 20 environments). The mean-field approximation is degenerate for two-agent settings.
• Seeds are arbitrary but fixed. The 7 seeds (99–105) are a design decision, not a sample from any population.

Reproducibility

Reproduce the paper's tables and figures from this dataset:

git clone https://github.com/vikpant/strategic-coopetition.git
cd strategic-coopetition
pip install -e ./coopetition_gym

# Point at the extracted dataset
python -m experiments.analyze all \
    --input-dir data/training/baseline_integrated/ \
    --output-dir data/analysis/

# Reward-type ablation comparison
python -m experiments.analyze reward-ablation \
    --input-baseline    data/training/baseline_integrated/ \
    --input-private     data/training/ablation_private/ \
    --input-cooperative data/training/ablation_cooperative/ \
    --output-dir        data/analysis/reward_ablation/

Regenerate the dataset from scratch (3,400 GPU-hours, approximately $8,100 USD on commodity cloud GPUs):

python -m experiments.campaign baseline --enable-checkpoints \
    --output data/training/baseline_integrated/
python -m experiments.campaign private --output data/training/ablation_private/
python -m experiments.campaign cooperative --output data/training/ablation_cooperative/

See REPRODUCE.md for full instructions.

Validation

Check dataset integrity after download:

python -m experiments.validate training data/training/

Expected output: 17,930 files, 62 expected NaN entries, 0 failed experiments.

Limitations

This dataset should not be used to:

• Train policies for deployment in actual business settings without extensive domain-specific validation. The validated case studies are abstract models, not operational authorizations.
• Claim results about empirical human cooperative behavior. The dataset reflects agent behavior under synthetic reward structures.

See the repository's DATASHEET.md for the complete Gebru et al. datasheet.

Citation

@article{pant2026coopetitiongym,
    title={Coopetition-Gym v1: A Formally Grounded Platform for Mixed-Motive
           Multi-Agent Reinforcement Learning under Strategic Coopetition},
    author={Pant, Vik and Yu, Eric},
    journal={arXiv preprint arXiv:2605.02063},
    year={2026}
}

@software{pant2026coopetitiongym_software,
    author={Pant, Vik and Yu, Eric},
    title={Coopetition-Gym: reproducibility package for the Coopetition-Gym benchmark},
    version={1.0.0},
    year={2026},
    publisher={Zenodo},
    doi={10.5281/zenodo.20015197}
}

Software archival deposit: https://doi.org/10.5281/zenodo.20015197 (concept DOI; resolves to latest version).

The benchmark environments are formalized in four foundational technical reports: arXiv:2510.18802 (TR-1), arXiv:2510.24909 (TR-2), arXiv:2601.16237 (TR-3), arXiv:2604.01240 (TR-4).

License

CC-BY-4.0 (Creative Commons Attribution 4.0 International). Users may share and adapt the dataset for any purpose, including commercial, provided attribution is given to the original authors.

Maintenance

• Issues and corrections: https://github.com/vikpant/strategic-coopetition/issues
• Contact: vik.pant@mail.utoronto.ca
• Changelog: https://github.com/vikpant/strategic-coopetition/blob/master/CHANGELOG.md

The v1 dataset is frozen for reproducibility. Future extensions will be released as versioned successors with their own datasheets.

Technical Reports

• TR-1: Computational Foundations for Strategic Coopetition: Formalizing Interdependence and Complementarity (arXiv:2510.18802)
• TR-2: Computational Foundations for Strategic Coopetition: Formalizing Trust and Reputation Dynamics (arXiv:2510.24909)
• TR-3: Computational Foundations for Strategic Coopetition: Formalizing Collective Action and Loyalty (arXiv:2601.16237)
• TR-4: Computational Foundations for Strategic Coopetition: Formalizing Sequential Interaction and Reciprocity (arXiv:2604.01240)