Datasetpublic

Bauxitiego/neural-mis-training-graphs

Synthetic Maximum Independent Set training graphs 500 synthetic Erdos-Renyi graphs (20-1,972 nodes, density 0.05-0.5), each labelled by 20 restarts of randomized-greedy construction plus (1,2)-exchange local search. Built to train Bauxitiego/neural-mis, a GCN evaluated against QOBLIB's Maximum Independent Set benchmark — code, evaluation, and honest results (including a documented failure) at github.com/Bauxitiego/neural-mis. Generated, not collected: free, unlimited… See the full description on the dataset page: https://huggingface.co/datasets/Bauxitiego/neural-mis-training-graphs.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

1likes37downloads

Dataset Card

Synthetic Maximum Independent Set training graphs

500 synthetic Erdos-Renyi graphs (20-1,972 nodes, density 0.05-0.5), each labelled by 20 restarts of randomized-greedy construction plus (1,2)-exchange local search. Built to train Bauxitiego/neural-mis, a GCN evaluated against QOBLIB's Maximum Independent Set benchmark — code, evaluation, and honest results (including a documented failure) at github.com/Bauxitiego/neural-mis.

Generated, not collected: free, unlimited, exactly-labelled by construction. Never overlaps QOBLIB's 50 evaluation instances — this dataset is training data only, the model's real evaluation happens entirely against QOBLIB's own ground truth.

Splits

train (450 graphs) and validation (50 graphs), a simple index split by generation order, not a reconstruction of the exact random subset used during actual training (that split came from a seeded PyTorch shuffle at training time; this one is simpler and explicit, documented as such rather than implying false precision).

Schema

Column	Type	Description
`graph_id`	int	Generation index, `[seed, graph_id]` was the RNG seed
`n`	int	Number of nodes
`density_p`	float	Erdos-Renyi edge probability used to generate the graph
`er_seed`	int	The specific seed passed to `networkx.erdos_renyi_graph`
`edge_src`, `edge_dst`	list\<int32\>	Edge list, one direction per edge, parallel arrays
`soft_label`	list\<float32\>	Per-node: fraction of near-best local-search restarts (within 1 of the best found) that included this node
`hard_label`	list\<int8\>	Per-node: 1 if in the single best independent set found across restarts, else 0
`best_size`	int	Size of the best independent set found for this graph

soft_label is the actual training target used (see the repo's src/dataset.py), not hard_label — MIS has many symmetric optima, and supervising on one arbitrary tie-break would teach a model to be confidently wrong about the others. hard_label is included for anyone who wants a single-solution view instead.

Reproduce from scratch

bash

python scripts/generate_training_data.py --n-graphs 500 --restarts 20 \
    --out-dir data/synthetic_train
python scripts/build_hf_dataset.py

License

Apache 2.0. Fully synthetic, no third-party data involved.