---
license: mit
pretty_name: remilp-data
---

# ReMILP dataset

MILP instances as bipartite variable-constraint graphs, with the labels of the
downstream tasks of the ReMILP paper.

| directory                        | instances                  | splits                                         | labels                    |
|----------------------------------|----------------------------|------------------------------------------------|---------------------------|
| `milp_evolve/`                   | 42,193                     | `pretrain/{train,val}`, `gap/{train,val,test}` | integrality gap           |
| `dmiplib/<CLASS>/<difficulty>/`  | 300 per pair (23 pairs)    | `train`, `val`, `test` = 200 / 50 / 50         | solution pool             |

Pairs, from [Distributional MIPLIB](https://huggingface.co/datasets/weiminhu/D-MIPLIB):
CA and IS easy/medium; SC, VC and GISP easy/medium/hard; MMCN medium-BI, hard-BI and
medium-BC; CFLP easy/medium; OTS easy/medium/hard; NNV easy; LB hard.

## Layout

```
<source>/manifest.json                {"source", "processing", "entries": [{"id", "split", "num_nodes", "num_edges"}, ...]}
<source>/<split>/graphs/<id>.pt.zst   a torch_geometric HeteroData, zstd-compressed
<source>/<split>/raw.tar              instances/<id>.(mps|lp).gz and labels/<id>.json.gz
```

## Labels

- `milp_evolve`: `lp_ip_gap = min(1, |IP - LP| / (max(|IP|, |LP|) + 1e-5))`.
- `dmiplib`: a pool of up to 300 solutions per instance (Gurobi, one hour per instance).
  `solution_probs` and `active_probs` are softmax-over-objective weighted averages over
  the pool of each binary variable's value and of each constraint's tightness.

## Use

```bash
hf download orailix/remilp-data --repo-type dataset --local-dir data --exclude "*/raw.tar"
```

## Attribution

The instances in this dataset come from [D-MIPLIB](https://huggingface.co/datasets/weiminhu/D-MIPLIB) and [MILP-Evolve](https://huggingface.co/datasets/microsoft/MILP-Evolve). All instances are re-distributed under their original licenses. Refer to `DATA_LICENSE.md` for more details.