Yuma-Ichikawa/qqa4co-bench

QQA4CO Combinatorial Optimization Benchmark Suite

A unified, pre-converted, ready-to-benchmark collection of Combinatorial-Optimization (CO) instances for discrete samplers, annealers, and learning-based solvers. Every family referenced in the PQQA paper Ichikawa & Iwashita, 2024 is reproduced here, together with a few broadly used community benchmarks (G-set, DIMACS COLOR, Edwards-Anderson). The dataset is designed to be solver-agnostic — a companion Python loader is shipped in QQA4CO, but every file is plain pickle(networkx.Graph) or numpy.savez, so it can be used from any framework (PyTorch, JAX, C++, Julia, ...).

Repository id history. The dataset was renamed from Yuma-Ichikawa/discs-co-bench → Yuma-Ichikawa/qqa4co-bench on 2026-04-20 to reflect the scope expansion beyond DISCS. Hugging Face preserves a redirect from the old id, but please update bookmarks, snapshots, and snapshot_download calls.

What's inside

The eight config_names (MaxCut, G-set, MIS, MaxClique, NormCut, Coloring, MIS-RRG, EA3D) collectively cover:

1. DISCS (NeurIPS 2023) — MaxCut / MIS / MaxClique / NormCut, repackaged from the original mixed pickle layouts into a uniform *.gpickle + manifest.jsonl format. We only repackage; always cite Goshvadi et al., 2023.
2. MaxCut G-set — the Helmberg & Rendl (2000) superset (G1..G67 + G70, G72, G77, G81, 71 graphs total) via Yinyu Ye's Stanford mirror. Best-known cuts track Benlic & Hao (2013) and Matsuda (2018). This is the de facto MaxCut benchmark for annealing and PI-GNN solvers.
3. PQQA reproduction set (arXiv:2409.02135v2) — full reproduction of every benchmark from §5.1–§5.5:
   • MIS on SATLIB (500 graphs) and Erdős–Rényi random graphs (ER-[700-800], ER-[9000-11000]) via the DISCS snapshot.
   • MIS on d-regular random graphs (RRGs) with d ∈ {20, 100} × n ∈ {10^4, 10^5, 10^6} — 5 seeds per cell, 30 instances total, ~9.7 GB.
   • Max Clique on the RB synthetic graphs and the SNAP Twitter real-world graph.
   • Max Cut on the DISCS ER/BA random-graph sizes and the Optsicom real-world set.
   • Balanced graph partition on VGG / MNIST-conv / ResNet / AlexNet / Inception-v3 computation graphs (re-uses the normcut/nets/ DISCS graphs with a different objective).
   • Graph Coloring on all 12 COLOR instances cited in Table 6 — Mycielski, square Queen, and the three DIMACS real-world graphs anna, jean, queen8_12.
4. 3D Edwards–Anderson spin glass — Gaussian and bimodal (±J) couplings on cubic lattices L ∈ {4, 6, 8}, 50 disorder realisations per cell, periodic boundary. A convenience extra for Ising / sampler benchmarks.

---

Coverage matrix vs. arXiv:2409.02135v2

Paper row	Instances	HF subset	num_instances	Best-known reference
§5.1 Table 1 — MIS SATLIB	500 CNFs, ≤1,347 nodes, ≤5,978 edges	mis/satlib/uf	500	KaMIS (Lamm et al., 2016; Hespe et al., 2019)
§5.1 Table 1 — MIS ER-[700-800]	128 ER graphs	mis/er/800	128	per-instance KaMIS (carried in manifest)
§5.1 Table 1 — MIS ER-[9000-11000]	16 ER graphs (see caveat ★)	mis/er/10k	16	KaMIS aggregate 381.31 (paper Table 5 footnote)
§5.1 Table 2 — MIS RRG d=20	n ∈ {10⁴, 10⁵, 10⁶}, 5 seeds each	mis-rrg/d20_n{10000,100000,1000000}	15	Barbier–Krzakała–Zdeborová 2013 RS density ρ = 0.2498
§5.1 Table 2 — MIS RRG d=100	n ∈ {10⁴, 10⁵, 10⁶}, 5 seeds each	mis-rrg/d100_n{10000,100000,1000000}	15	Barbier–Krzakała–Zdeborová 2013 RS density ρ = 0.0669
§5.2 Table 3 — Max Clique RB	Xu et al. (2007) RBtest	maxclique/rb/all	500	per-instance DISCS reference
§5.2 Table 3 — Max Clique Twitter	SNAP Twitter ego-network	maxclique/twitter/all	196	per-instance DISCS reference
§5.3 Fig. 3 — Max Cut ER	7 size buckets, 16–1,100 nodes	maxcut/er/er-0.15-n-*	700	Gurobi 1 h (DISCS)
§5.3 Fig. 3 — Max Cut BA	7 size buckets, 16–1,100 nodes	maxcut/ba/ba-4-n-*	700	Gurobi 1 h (DISCS)
§5.3 Table 4 — Max Cut Optsicom	10 real-world ±1/0 graphs	maxcut/optsicom/b	10	per-instance DISCS reference
§5.4 Table 5 — Balanced partition	VGG, MNIST-conv, ResNet, AlexNet, Inception-v3	normcut/nets/{VGG,MNIST,RESNET,ALEXNET,INCEPTION} (loader swaps the objective)	5	no reference optimum; comparative only
§5.5 Table 6 — Coloring (Myciel)	myciel5, myciel6 (plus {3,4,7} as extras)	coloring/myciel	5	chromatic number (Mycielski 1955)
§5.5 Table 6 — Coloring (Queen)	queen{5..13}_{5..13} (7 of 9 used in Table 6)	coloring/queen	9	tabulated chromatic number
§5.5 Table 6 — Coloring (DIMACS)	anna, jean, queen8_12	coloring/dimacs	3	chromatic number from Trick, 2002

★ ER-[9000-11000] caveat. The 16 ER_9000_11000_*.gpickle graphs are the exact instances released with DISCS and used by PQQA. The upstream DISCS conversion does not include per-instance KaMIS labels, so the manifest records best_known: null and qqa bench-run reports ApR = NaN on this subset. The paper quotes the aggregate KaMIS average (381.31, Table 5 footnote) — reproducing the Table 1 ApR numbers requires either running KaMIS yourself on each of the 16 graphs or dividing raw IS sizes by that aggregate. We are tracking upstream to restore per-instance labels.

Convenience extras shipped here that are not in the paper: gset/standard (G-set MaxCut benchmark), mis/er_density/* (ER MIS density sweeps), normcut/nets/{BABELFISH,NMT,TTS} (additional DISCS compute graphs beyond §5.4), and ea3d/* (3D Ising spin glass).

---

Quick start

Option A — Via QQA4CO (recommended, three one-liners)

git clone https://github.com/Yuma-Ichikawa/QQA4CO.git
cd QQA4CO
pip install -e ".[discs,dev]"

make bench-all-setup                                   # pull every family from this dataset (~14 GB)
qqa bench-run --suite all --output bench_results/mine.json
qqa bench-plot bench_results/mine.json --output report.png

qqa bench-plot renders a publication-quality 2×2 figure with per-subset bars, a radar chart, feasibility ratios, and per-instance violin+strip plots. Pass multiple JSON files to produce an A/B/C comparison:

qqa bench-plot bench_results/baseline.json bench_results/mine.json \
    --labels "baseline" "my method" \
    --title "ablation vs. baseline" \
    --output ab.png

Scoping to a single family or subset:

qqa bench-list                                         # show every available suite
qqa bench-run --suite gset                      --instances 5
qqa bench-run --suite coloring-dimacs           --instances 3    # anna, jean, queen8_12
qqa bench-run --suite mis-rrg-d20_n10000        --instances 5    # PQQA Table 2, n=10^4
qqa bench-run --suite mis-rrg-d100_n1000000     --instances 1    # PQQA Table 2, n=10^6
qqa bench-run --suite ea3d-gaussian-L4          --instances 3
qqa bench-run --suite balanced-partition-nets-INCEPTION --instances 1

The same three verbs are available from Python:

from qqa import bench

bench.list_suites()                                       # {suite_id: (family, graph_type, subset)}
bench.run("gset", instances=5, output="mine.json")        # writes to ./bench_results/
bench.plot(["bench_results/mine.json"], output="report.png")

Option B — Plain Python (no QQA4CO required)

from huggingface_hub import snapshot_download
import json, pickle, pathlib
import networkx as nx
import numpy as np

local = snapshot_download(
    repo_id="Yuma-Ichikawa/qqa4co-bench",
    repo_type="dataset",
    allow_patterns=["gset/**"],       # ~30 MB; omit to download everything
)
root = pathlib.Path(local)

for line in (root / "gset/standard/manifest.jsonl").open():
    rec = json.loads(line)
    with (root / "gset/standard" / rec["file"]).open("rb") as fh:
        g: nx.Graph = pickle.load(fh)
    print(rec["id"], "n=", g.number_of_nodes(), "best_known=", rec["best_known"])

# 3D Edwards-Anderson: coupling lists shipped as .npz
ea = np.load(root / "ea3d/gaussian/L4/0001.npz")
print("L=", int(ea["L"]), "num_couplings=", len(ea["J"]))

Option C — datasets library

Each of the eight families is a separate config_name, so you can stream just the subset you care about:

from datasets import load_dataset

ds = load_dataset(
    "Yuma-Ichikawa/qqa4co-bench",
    name="mis-rrg",                  # or maxcut, gset, mis, maxclique, ...
    split="train",
    streaming=True,
)
for rec in ds.take(1):
    print(rec.keys())               # -> dict_keys(['path', 'bytes', ...])

---

Layout

.
├── maxcut/                              (~3.3 GB, ~9,000 instances; DISCS)
│   ├── ba/ba-4-n-{16-20,32-40,64-75,128-150,256-300,512-600,1024-1100}/
│   ├── er/er-0.15-n-{16-20,32-40,64-75,128-150,256-300,512-600,1024-1100}/
│   └── optsicom/b/                  (10 real-world Optsicom graphs)
├── mis/                                 (~365 MB; DISCS)
│   ├── satlib/uf/                   (500 SATLIB CNF → IS graph encodings)
│   ├── er/{800,10k}/                (ER-[700-800] × 128, ER-[9000-11000] × 16)
│   └── er_density/{0.05,0.10,0.20,0.25}/  (additional ER density sweeps)
├── maxclique/                           (~131 MB; DISCS)
│   ├── rb/all/                      (Xu et al. 2007 RBtest, 500 instances)
│   └── twitter/all/                 (SNAP Twitter ego-network, 196 instances)
├── normcut/                             (~7.4 MB; DISCS; also consumed by Balanced Partition)
│   └── nets/{VGG,MNIST,RESNET,ALEXNET,INCEPTION,BABELFISH,NMT,TTS}/
├── gset/                                (~30 MB; 71 G-set graphs)
│   └── standard/ (G1..G67, G70, G72, G77, G81)
├── coloring/                            (~350 KB; procedural + DIMACS)
│   ├── myciel/     (Mycielski graphs k=3..7; chromatic number = k)
│   ├── queen/      (queen-attack graphs on k × k boards, k = 5..13)
│   └── dimacs/     (DIMACS COLOR real-world: anna, jean, queen8_12 — arXiv Table 6)
├── mis-rrg/                             (~9.7 GB; procedural)
│   ├── d20_n10000/    (d=20, n=10^4, 5 seeds; PQQA §5.1 Table 2)
│   ├── d20_n100000/   (d=20, n=10^5)
│   ├── d20_n1000000/  (d=20, n=10^6; ~1.6 GB total)
│   ├── d100_n10000/   (d=100, n=10^4; dense regime)
│   ├── d100_n100000/  (d=100, n=10^5)
│   └── d100_n1000000/ (d=100, n=10^6; ~8 GB total)
└── ea3d/                                (~280 KB; procedural)
    ├── gaussian/{L4,L6,L8}/    (N(0,1) couplings; cubic lattice, PBC)
    └── bimodal/{L4,L6,L8}/     (±1 couplings; ±J spin glass)

Per-instance file formats

• *.gpickle (DISCS, G-set, Coloring, MIS-RRG) — pickle.dump(networkx.Graph). Edge weights are carried on the weight edge attribute (±1 for the G-set ±1 families, real-valued for the Gaussian-weighted families).
• *.npz (EA3D) — sparse coupling list with arrays i, j, J, L (lattice edge list + cube side). The QQA4CO loader qqa.datasets.ea3d reassembles the J matrix and instantiates qqa.problems.EdwardsAnderson.
• manifest.jsonl — one JSON object per line, with at least {id, file, best_known, source}. Family-specific extras: num_colors, best_known_source (coloring); d, n, seed (mis-rrg); num_spins, L, distribution (ea3d); num_nodes, num_edges, best_known_source, source_url (gset); problem, graph_type, subset, source (DISCS).

---

Best-known references

Family	best_known semantics
maxcut / mis / maxclique	upstream DISCS reference, higher is better
gset	Benlic & Hao (2013), Matsuda (2018); higher is better
normcut	upstream DISCS reference, lower is better
coloring	0 (= minimum number of edge conflicts a proper K-colouring must reach); num_colors carries the (known) chromatic number
mis-rrg	Barbier–Krzakała–Zdeborová (2013) replica-symmetric asymptotic MIS density × n (ρ_{d=20} = 0.2498, ρ_{d=100} = 0.0669)
ea3d	brute-force ground-state energy for N ≤ 20; NaN for larger lattices
balanced-partition (on normcut/nets/*)	no published reference (NaN); use for comparative runs only

Approximation Ratio (ApR) conventions — consistent with the PQQA paper and with qqa bench-plot:

• Maximization (MaxCut, MIS, MaxClique): ApR = value / best_known, so ApR ≤ 1 at optimality.
• Minimization (NormCut, Coloring objective-as-conflicts): ApR = best_known / value, so ApR ≤ 1 at optimality.
• Subsets whose best_known is null or NaN report ApR = NaN and are excluded from aggregate statistics.

---

Caveats and reproducibility notes

• G-set has upstream holes. G68, G69, G71, G73–G76, G78–G80 return HTTP 404 from the Stanford mirror and are therefore not hosted here. scripts/fetch_gset_data.py logs the skipped indices.
• normcut/nets/ graphs are highly disconnected. Several computation graphs (BABELFISH, TTS, ALEXNET, VGG, …) consist of a giant component plus dozens of 2–4 node fragments. A naive solver reaches Ncut = 0 trivially by isolating the small components. Restrict to the largest connected component when you want a non-trivial bisection.
• TRANSFORMER.pkl was empty in the upstream DISCS release — omitted.
• normcut-gap_rand is not present in the upstream DISCS tarball — omitted.
• MIS on RRG at n = 10^6 is fully hosted (5 seeds × 2 degrees, ~9.7 GB total). Each d=100, n=10^6 .gpickle is ~1.6 GB on disk; the full adjacency matrix does not fit in a single dense tensor (n² = 10¹² entries), so solvers must use sparse representations. Re-generate locally with python scripts/generate_rrg_instances.py --include-huge if needed.
• Barbier d=100 density correction (Apr 2026). Earlier snapshots of scripts/generate_rrg_instances.py carried _BARBIER_DENSITY[100] = 0.1360 (a typo; that is the d=20 asymptotic density), which inflated best_known by roughly 2× on the d100_n10000 manifest. The current version uses the correct ρ_{d=100} = 0.0669; all d100_n* manifests have been rebuilt.
• DIMACS coloring/dimacs/ was added 2026-04-22. The three real-world instances anna, jean, queen8_12 (Table 6 of the PQQA paper) were previously missing. They are now fetched once from Trick's canonical mirror by scripts/generate_coloring_instances.py, cached locally under data/coloring/_dimacs_cache/, and shipped on the Hub.

---

Sources

Resource	URL
PQQA paper	https://arxiv.org/abs/2409.02135
DISCS paper	https://openreview.net/forum?id=oi1MUMk5NF
DISCS upstream code & data	https://github.com/google-research/discs
G-set Stanford mirror	https://web.stanford.edu/~yyye/yyye/Gset/
DIMACS COLOR instances (Trick)	https://mat.tepper.cmu.edu/COLOR/instances
SATLIB	https://www.cs.ubc.ca/~hoos/SATLIB/benchm.html
SNAP Twitter graph	https://snap.stanford.edu/data/ego-Twitter.html
KaMIS	https://karlsruhemis.github.io/
Conversion / generator scripts	https://github.com/Yuma-Ichikawa/QQA4CO/tree/main/scripts
QQA4CO per-family READMEs	https://github.com/Yuma-Ichikawa/QQA4CO/tree/main/data
This dataset's versioned README	this file

---

Citation

If you use this dataset, please cite the PQQA paper (the design target for the scope and the paper that tabulates the expected ApR numbers):

@article{ichikawa2024pqqa,
  title   = {Optimization by Parallel Quasi-Quantum Annealing with
             Gradient-Based Sampling},
  author  = {Ichikawa, Yuma and Iwashita, Hiroshi},
  journal = {arXiv preprint arXiv:2409.02135},
  year    = {2024},
  url     = {https://arxiv.org/abs/2409.02135}
}

In addition, please cite the upstream source(s) of whichever subset(s) you use.

DISCS subsets (maxcut, mis, maxclique, normcut) — data is Goshvadi et al.'s; we only repackage:

@inproceedings{goshvadi2023discs,
  title     = {{DISCS}: A Benchmark for Discrete Sampling},
  author    = {Goshvadi, Katayoon and Sun, Haoran and Liu, Xingchao
               and Nova, Azade and Zhang, Ruqi and Grathwohl, Will
               and Schuurmans, Dale and Dai, Hanjun},
  booktitle = {Advances in Neural Information Processing Systems
               (NeurIPS Datasets and Benchmarks Track)},
  year      = {2023},
  url       = {https://openreview.net/forum?id=oi1MUMk5NF}
}

@inproceedings{sun2023revisiting,
  title     = {Revisiting Sampling for Combinatorial Optimization},
  author    = {Sun, Haoran and Goshvadi, Katayoon and Nova, Azade and
               Schuurmans, Dale and Dai, Hanjun},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2023}
}

MIS on SATLIB — cite the original SATLIB benchmark:

@incollection{hoos2000satlib,
  title     = {{SATLIB}: An Online Resource for Research on {SAT}},
  author    = {Hoos, Holger H. and St{\"u}tzle, Thomas},
  booktitle = {SAT 2000: Highlights of Satisfiability Research in the Year 2000},
  editor    = {Gent, I. P. and van Maaren, H. and Walsh, T.},
  publisher = {IOS Press},
  pages     = {283--292},
  year      = {2000}
}

MIS reference solutions (KaMIS) — used for SATLIB and ER:

@inproceedings{lamm2016finding,
  title     = {Finding Near-Optimal Independent Sets at Scale},
  author    = {Lamm, Sebastian and Sanders, Peter and Schulz,
               Christian and Strash, Darren and Werneck, Renato F.},
  booktitle = {Proceedings of the 18th Meeting on Algorithm Engineering
               and Experiments (ALENEX)},
  year      = {2016},
  doi       = {10.1137/1.9781611974317.12}
}

@inproceedings{hespe2019scalable,
  title     = {Scalable Kernelization for Maximum Independent Sets},
  author    = {Hespe, Demian and Schulz, Christian and Strash, Darren},
  booktitle = {ACM Journal of Experimental Algorithmics (JEA)},
  year      = {2019},
  doi       = {10.1145/3355502}
}

Max Clique — RB random model and SNAP Twitter ego-graph:

@article{xu2007random,
  title   = {Random Constraint Satisfaction: Easy Generation of Hard
             (Satisfiable) Instances},
  author  = {Xu, Ke and Boussemart, Fr{\'e}d{\'e}ric and Hemery, Fred
             and Lecoutre, Christophe},
  journal = {Artificial Intelligence},
  volume  = {171},
  number  = {8-9},
  pages   = {514--534},
  year    = {2007},
  doi     = {10.1016/j.artint.2007.04.001}
}

@misc{leskovec2014snap,
  title  = {{SNAP} Datasets: {Stanford} Large Network Dataset Collection},
  author = {Leskovec, Jure and Krevl, Andrej},
  year   = {2014},
  url    = {http://snap.stanford.edu/data}
}

Max Cut G-set — seminal spectral-bundle paper and best-known reference:

@article{helmberg2000gset,
  title   = {A Spectral Bundle Method for Semidefinite Programming},
  author  = {Helmberg, Christoph and Rendl, Franz},
  journal = {SIAM Journal on Optimization},
  volume  = {10},
  number  = {3},
  pages   = {673--696},
  year    = {2000},
  doi     = {10.1137/S1052623497328987}
}

@article{benlic2013bls,
  title   = {Breakout Local Search for the Max-Cut Problem},
  author  = {Benlic, Una and Hao, Jin-Kao},
  journal = {Engineering Applications of Artificial Intelligence},
  volume  = {26},
  number  = {3},
  pages   = {1162--1173},
  year    = {2013},
  doi     = {10.1016/j.engappai.2012.09.001}
}

@article{matsuda2018gset,
  title   = {{MQLib}: Infrastructure for Empirical Evaluation of Heuristics
             for Max-Cut and {QUBO}},
  author  = {Dunning, Iain and Gupta, Swati and Silberholz, John},
  journal = {INFORMS Journal on Computing},
  volume  = {30},
  number  = {3},
  pages   = {608--624},
  year    = {2018},
  doi     = {10.1287/ijoc.2017.0798}
}

MIS on regular random graphs — asymptotic density (used as best_known) and the original hardness argument:

@article{barbier2013hard,
  title   = {The Hard-Core Model on Random Graphs Revisited},
  author  = {Barbier, Jean and Krzakala, Florent and Zdeborov{\'a},
             Lenka and Zhang, Pan},
  journal = {Journal of Physics: Conference Series},
  volume  = {473},
  pages   = {012021},
  year    = {2013},
  doi     = {10.1088/1742-6596/473/1/012021}
}

@article{angelini2023modern,
  title   = {Modern Graph Neural Networks Do Worse than Classical
             Greedy Algorithms in Solving Combinatorial Optimization
             Problems like Maximum Independent Set},
  author  = {Angelini, Maria Chiara and Ricci-Tersenghi, Federico},
  journal = {Nature Machine Intelligence},
  volume  = {5},
  pages   = {29--31},
  year    = {2023},
  doi     = {10.1038/s42256-022-00589-y}
}

Graph Coloring (COLOR / DIMACS) — Trick's canonical compilation and the two procedural families that cover most of Table 6:

@misc{trick2002color,
  title  = {Graph Coloring Instances},
  author = {Trick, Michael A.},
  year   = {2002},
  note   = {Carnegie Mellon, \url{https://mat.tepper.cmu.edu/COLOR/instances.html}}
}

@article{mycielski1955coloring,
  title   = {Sur le coloriage des graphes},
  author  = {Mycielski, Jan},
  journal = {Colloquium Mathematicae},
  volume  = {3},
  number  = {2},
  pages   = {161--162},
  year    = {1955}
}

@book{knuth1993sgb,
  title     = {The Stanford {GraphBase}: A Platform for Combinatorial Computing},
  author    = {Knuth, Donald E.},
  publisher = {ACM Press / Addison-Wesley},
  year      = {1993}
}

Balanced Graph Partition — GAP baseline and hMETIS framework:

@inproceedings{nazi2019gap,
  title     = {{GAP}: Generalizable Approximate Graph Partitioning Framework},
  author    = {Nazi, Azade and Hang, Will and Goldie, Anna and
               Ravi, Sujith and Mirhoseini, Azalia},
  booktitle = {ICLR Workshop on Representation Learning on Graphs and Manifolds},
  year      = {2019}
}

@article{karypis1999multilevel,
  title   = {Multilevel Hypergraph Partitioning: Applications in {VLSI} Domain},
  author  = {Karypis, George and Aggarwal, Rajat and Kumar, Vipin
             and Shekhar, Shashi},
  journal = {IEEE Transactions on Very Large Scale Integration (VLSI) Systems},
  volume  = {7},
  number  = {1},
  pages   = {69--79},
  year    = {1999},
  doi     = {10.1109/92.748202}
}

Optionally, also cite the QQA4CO repackaging infrastructure:

@misc{ichikawa2026qqa4co,
  title  = {{QQA4CO}: A Reproducible GPU Benchmark Suite for
            Combinatorial Optimization},
  author = {Ichikawa, Yuma},
  year   = {2026},
  url    = {https://github.com/Yuma-Ichikawa/QQA4CO}
}

---

License

This repackaging is released under Apache-2.0. The underlying DISCS, SATLIB, DIMACS, SNAP, and G-set instances inherit the licenses of their original sources (see the Sources table above); please consult those upstream links if you redistribute. The procedurally generated coloring/{myciel,queen}, mis-rrg, and ea3d subsets are original to this dataset and are released under Apache-2.0.

---

How to benchmark a new solver

A detailed guide (Python + CLI + Make, with ratio conventions and per-family feasibility definitions) lives in the QQA4CO docs: docs/how-to/benchmark.md.

Changelog

• 2026-04-22 — added coloring/dimacs/ (anna, jean, queen8_12); arXiv-2409.02135v2 Table 6 coverage is now 12/12.
• 2026-04-20 — added the d20_n{10⁵, 10⁶} and d100_n{10⁵, 10⁶} RRG cells (four 5-seed subsets, ~9 GB); corrected the Barbier d=100 density bug on d100_n10000/manifest.jsonl (best_known ×½).
• 2026-04-20 — repository renamed from discs-co-bench → qqa4co-bench.