---
title: README
emoji: 🌍
colorFrom: indigo
colorTo: red
sdk: static
pinned: false
---

# PGLearn Datasets

This HuggingFace organization hosts the PGLearn datasets, described in https://arxiv.org/abs/2505.22825.

| Collection                   | Link                                                          | #feasible | #samples |
|------------------------------|---------------------------------------------------------------|-----------|----------|
| Small (#buses ≤ 1000)        | https://huggingface.co/collections/PGLearn/pglearn-small      | ~5.749M   | 6M       |
| Medium (#buses ≤ 5000)       | https://huggingface.co/collections/PGLearn/pglearn-medium     | ~1.573M   | 1.75M    |
| Large (#buses ≤ 10000)       | https://huggingface.co/collections/PGLearn/pglearn-large      | ~253.6K   | 300K     |
| Extra-Large (#buses > 10000) | https://huggingface.co/collections/PGLearn/pglearn-extralarge | ~69.9K    | 75K      |
| N-1 contingency cases        | https://huggingface.co/collections/PGLearn/pglearn-n-1        | ~3.575M   | 4.2M     |

## Citation

```bibtex
@article{klamkin2025pglearn,
  title={PGLearn--An Open-Source Learning Toolkit for Optimal Power Flow},
  author={Klamkin, Michael and Tanneau, Mathieu and Van Hentenryck, Pascal},
  journal={arXiv preprint arXiv:2505.22825},
  year={2025}
}
```

## Instructions

  <details> <summary>Click to open usage instructions. </summary> 

The datasets are available in two formats: Parquet and HDF5. 

- Parquet: Use the HuggingFace [datasets](https://github.com/huggingface/datasets) package as usual, see their documentation for further instructions.
- HDF5: Use the `snapshot_download` function from [huggingface_hub](https://github.com/huggingface/huggingface_hub):
```python
from huggingface_hub import snapshot_download

snapshot_download(
  "PGLearn/PGLearn-Small-14_ieee",
  repo_type="dataset",
  local_dir="./14_ieee",    # where to put it
  revision="script",        # IMPORTANT: grab the HDF5 files, not the parquet files

  # you can set filters, e.g. if we only want the DC samples:
  allow_patterns=[
    "*/DCOPF/*", "*input*", "case.json.gz", "config.toml",
  ],
  ignore_patterns=[
    "infeasible/*"
  ],
)
```

Then, you can load them with `h5py`. Note that for some large cases, the dual solution data had to be split up into multiple files. The below helper can decompress and reconstruct these files:
```python
from pathlib import Path
import gzip, shutil
def open_maybe_gzip_cat(path: str | list):
    if isinstance(path, list):
        dest = Path(path[0]).parent.with_suffix(".h5")
        if not dest.exists():
            with open(dest, "wb") as dest_f:
                for piece in path:
                    with open(piece, "rb") as piece_f:
                        shutil.copyfileobj(piece_f, dest_f)
            shutil.rmtree(Path(piece).parent)
        path = dest.as_posix()
    return gzip.open(path, "rb") if path.endswith(".gz") else open(path, "rb")

primal = h5py.File(open_maybe_gzip_cat("data/SOCOPF/primal.h5.gz"))
dual = h5py.File(open_maybe_gzip_cat(
  ["data/SOCOPF/dual/xaa","data/SOCOPF/dual/xab","data/SOCOPF/dual/xac"]
), "r")
```

If using the HDF5 dataset more than once, it is recommended to pre-decompress the files. The following helper can do this:
```python
from pathlib import Path
import gzip, shutil
for src in Path("./14_ieee").rglob("*.h5.gz"):
  dest = src.with_suffix("")
  with gzip.open(src, "rb") as fsrc, open(dest, "wb") as fdest:
    shutil.copyfileobj(fsrc, fdest)
    src.unlink() # optional; delete the compressed files
```

</details>