README.md

---
license: mit
task_categories:
- reinforcement-learning
tags:
- inventory-management
- supply-chain
- reinforcement-learning
- operations-research
- gymnasium
pretty_name: gym-invmgmt trained agent weights
---

# gym-invmgmt trained agent weights

This dataset repository hosts the trained agent checkpoints and normalization statistics for the `gym-invmgmt` benchmark release.

Contents are stored under `data/models/` and are intended to be downloaded into the same path in the GitHub repository.

## Related repositories

- Paper: [arXiv:2605.11355](https://arxiv.org/abs/2605.11355)
- PyPI package: [gym-invmgmt](https://pypi.org/project/gym-invmgmt/)
- Final paper/code repository: [r2barati/gym-invmgmt-paper](https://github.com/r2barati/gym-invmgmt-paper)
- Standalone Gymnasium environment package: [r2barati/gym-invmgmt](https://github.com/r2barati/gym-invmgmt)

## Included

- Stable-Baselines3 PPO/SAC checkpoints (`.zip`)
- Imitation-learning / DAgger / GNN-IL checkpoints (`.zip`)
- Matching VecNormalize statistics (`.pkl`)
- `models_manifest.json` with SHA-256 checksums

## Not included

- Third-party Qwen GGUF LLM weights are not re-hosted here. Use the original model source or place the expected file at `data/models/qwen2.5-1.5b-instruct-q4_k_m.gguf` for optional LLM reruns.
- Public retail datasets are not re-hosted here. Use the repository download script and source-specific licenses.

## Download

From the GitHub repository root:

```bash
bash download_weights.sh
```

or with Python:

```python
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='rezabarati/gym-invmgmt-weights',
    repo_type='dataset',
    local_dir='.',
    allow_patterns=['data/models/*', 'models_manifest.json'],
)
```

## Citation

If you use these trained checkpoints, please cite:

```bibtex
@misc{barati2026gyminvmgmt,
  title = {gym-invmgmt: An Open Benchmarking Framework for Inventory Management Methods},
  author = {Barati, Reza and Hu, Qinmin Vivian},
  year = {2026},
  eprint = {2605.11355},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG}
}
```