conashin/hgp-benchmark-unweighted

Hypergraph Partitioning Benchmark — Unweighted Netlist Format

Converted benchmark dataset for two-way hypergraph partitioning (min-cut under balance constraint).
Derived from the integrated benchmark collection at Zenodo 15373339, re-formatted into a plain unweighted netlist format suitable for EDA partitioning toolchains.

---

Dataset Summary

Property	Value
Total cases	489
Cell count range	5,804 – 9,128,020
Net count range	163 – 6,920,306
Pin count range	6,323 – 57,156,537
Average degree range	1.45 – 1,825
Original format	hMETIS .hgr
Converted format	Unweighted netlist (see below)
License	CC-BY-4.0

---

Scale Tiers

Cases are grouped into four tiers by total pin count (pins = sum of net sizes = cells × avg_degree):

Tier	Pin count	Cases
tiny	< 100K	29
small	100K – 500K	137
medium	500K – 5M	221
large	> 5M	102

Directory layout after classification:

data/
├── tiny/      # e.g. case153, case47
├── small/     # e.g. case100, case108
├── medium/    # e.g. case112, case113
└── large/     # e.g. case117, case488

---

File Format

Each case is a plain-text file. One net per line:

NET <net_id> { <cell_id> <cell_id> ... }

• <net_id> : string identifier, e.g. n1, n42
• <cell_id> : string identifier, e.g. c1, c200
• All weights are unit (unweighted); edge and vertex weights from the original .hgr files are discarded
• Cells that appear in no net are omitted (isolated vertices)

Example (case1, 10 cells, 8 nets)

NET n1 { c2 c6 }
NET n2 { c1 c7 c9 c10 }
NET n3 { c5 c9 }
NET n4 { c2 c6 c9 }
NET n5 { c5 c6 }
NET n6 { c3 c5 c6 c8 c10 }
NET n7 { c4 c6 }
NET n8 { c4 c5 c6 c7 c8 }

---

Partitioning Problem Definition

Given a hypergraph H = (V, E), find a balanced bipartition (A, B) that minimises cut size:

minimise   |{ e ∈ E : e ∩ A ≠ ∅ and e ∩ B ≠ ∅ }|
subject to |A| - |B|  ≤  |V| / 5

The balance constraint allows at most 20% imbalance between partition sizes.

---

Conversion

Original .hgr files (hMETIS format) were converted using the script below. Edge weights and vertex weights were stripped; the net-cell incidence structure is preserved exactly.

python hgr_to_pa3.py input.hgr output.case

Conversion script: hgr_to_pa3.py
Key flags: --zero-based for 0-indexed vertex IDs, --add-isolated-singleton to retain cells that appear in no net.

---

Usage Example (Python)

def parse_case(filepath):
    """Parse a case file into a list of nets (each net = list of cell names)."""
    nets = []
    with open(filepath) as f:
        for line in f:
            line = line.strip()
            if not line.startswith("NET"):
                continue
            # NET n1 { c2 c6 }
            inner = line[line.index('{') + 1 : line.index('}')]
            cells = inner.split()
            nets.append(cells)
    return nets

nets = parse_case("data/tiny/case1")
n_nets  = len(nets)
n_cells = len({c for net in nets for c in net})
n_pins  = sum(len(net) for net in nets)
print(f"nets={n_nets}, cells={n_cells}, pins={n_pins}")
# nets=8, cells=10, pins=22

---

Download

Download a single tier (recommended)

Python — huggingface_hub (downloads raw files, no parsing):

from huggingface_hub import snapshot_download

# Download only the tiny tier (~seconds)
local_dir = snapshot_download(
    repo_id="conashin/hgp-benchmark-unweighted",
    repo_type="dataset",
    allow_patterns="data/tiny/*",
    local_dir="./hgp_data",
)

Shell — huggingface-cli:

huggingface-cli download Yconashin/hgp-benchmark-unweighted \
    --repo-type dataset \
    --include "data/tiny/*" \
    --local-dir ./hgp_data

Download via datasets library (config-based)

from datasets import load_dataset

# Available configs: tiny | small | medium | large | all
ds = load_dataset(
    "conashin/hgp-benchmark-unweighted",
    "small",         # change to the tier you need
    split="train",
)

Note: load_dataset returns file metadata objects. Use the huggingface_hub approach above if you need the raw netlist files directly.

Download all tiers

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="conashin/hgp-benchmark-unweighted",
    repo_type="dataset",
    local_dir="./hgp_data",
)

# Shell equivalent
huggingface-cli download conashin/hgp-benchmark-unweighted \
    --repo-type dataset \
    --local-dir ./hgp_data

---

Dataset Sources

This dataset is a derived work converted from the following original benchmark collections, all published under CC-BY-4.0:

Source	DOI	Description
Compiled benchmark	10.5281/zenodo.15373339	Integrated collection (primary source)
KaHyPar benchmark	10.5281/zenodo.291466	ISPD98, DAC2012, SAT2014, UF-SPM
Scalable HGP benchmark	10.5281/zenodo.15386567	set_A_M_HG, set_B_L_HG
Memetic HGP benchmark	10.5281/zenodo.15387992	100 hypergraphs

Original benchmark instances derive from:

• ISPD98 Circuit Benchmark Suite
• DAC 2012 Routability-Driven Placement Contest
• SAT Competition 2014 (application track)
• University of Florida Sparse Matrix Collection

---

Citation

If you use this dataset, please cite the original sources:

@dataset{schlag2017benchmark,
  author    = {Schlag, Sebastian},
  title     = {A Benchmark Set for Multilevel Hypergraph Partitioning Algorithms},
  year      = {2017},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.291466},
  url       = {https://zenodo.org/records/291466}
}

@dataset{zenodo15373339,
  title     = {Hypergraph Partitioning Benchmark (Integrated Collection)},
  year      = {2025},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.15373339},
  url       = {https://zenodo.org/records/15373339}
}

---

License

This derived dataset is released under Creative Commons Attribution 4.0 International (CC-BY-4.0), consistent with the license of the original source material.

You are free to share and adapt this dataset for any purpose, provided that appropriate credit is given to the original authors and a link to the license is included.