---
pretty_name: SGPL local benchmark graphs
language:
- en
tags:
- graphs
- graph-algorithms
- benchmarking
- pagerank
- shortest-paths
---
# SGPL local benchmark graphs

The **61-graph local corpus** used for BFS, SSSP, connected components, and PageRank benchmark campaigns. This is the exact input corpus used in the local experiments through 24 October 2026; it is not the complete upstream graph collections.

Each graph is a separate `.tar.gz` archive. `manifest.json` records its original graph identifier, vertex/arc counts, direction settings, source vertex, synthetic-weight provenance, and SHA-256 checksums for the archive and every graph file. File bytes are preserved from the frozen benchmark inputs. No benchmark is run when packaging these files.

## Files in each graph archive

- `original.edges`: original benchmark edge-list input; integer vertex IDs, retaining its original occurrence order and duplicate edges.
- `graph.bin`: exact prepared binary CSR input used by the local graph benchmark harness. This is a benchmark-specific binary representation, not an Arrow or Parquet dataset. Its header is 64 bytes; vertex count and arc count are little-endian uint64 values at byte offsets 16 and 24. They are followed by `n` little-endian uint64 row offsets and `m` little-endian uint32 destination IDs. The terminal row offset is the header's `m`.
- `weighted-original.wel`: original edge-list occurrences with the benchmark's synthetic SSSP weights added as a third column.
- `weights.f32`: little-endian float32 synthetic weights in prepared CSR arc order, used by the saved Graptor comparison.
- `weights.i32`: the same synthetic weights as little-endian signed int32 in prepared CSR arc order.

The original edge list and prepared CSR need not have the same row count. Use the manifest's direction/storage fields: undirected inputs are represented with both arc directions in CSR. Preserve duplicate edges and sparse vertex IDs when reproducing the benchmarks.

## Synthetic SSSP weights

Weights are **synthetic positive integers from 1 through 255**, generated with seed **2358**. They must not be described as original or real-world edge weights. The two weight sidecars have the same numeric values and differ in encoding. BFS, CC, and PageRank use the unweighted topology.

## Download and extract one graph

```python
from huggingface_hub import hf_hub_download
import tarfile

archive = hf_hub_download(repo_id="ksahib/graphs",
                          filename="dimacs__hamming8-2.tar.gz", repo_type="dataset")
with tarfile.open(archive, "r:gz") as graph_archive:
    graph_archive.extractall("graphs", filter="data")
```

## Integrity and provenance

`SHA256SUMS` covers the distributed archives and manifest. Each archive contains only the five graph-input files listed above. The package excludes source code, binaries, logs, credentials, private machine paths, and performance reports.

The graph identifiers and collection group names are retained from the frozen benchmark manifest. That manifest does not establish a corpus-wide redistribution license or complete original download URLs; no new license is asserted here. Consult the relevant upstream graph collection's terms before reuse.