NunoBatista/gflowqec-data

GFlowQEC data

The large data files behind GFlowQEC: Generative Flow Networks that Learn to Sample Quantum LDPC Codes. The code, tests and small files are in the GitHub repository GFlowQEC. Paths match the repository, so this dataset unpacked into the repository root restores data/.

Paper: arXiv:2611.06819.

Path	Size	Contents
data/n180/	816 MB, 700 files	Search runs, GFlowQEC-offline and matched-screen samples, screen and certification files, and the reference bounds at n = 180, with ten seeds (0-9) for the main deployment comparisons.
data/n300/	804 MB, 552 files	Search runs at n = 300 with ten seeds (0-5 and 7-10, seed 6 was the tuning seed), their re-estimates on fresh shots, files of uniform draws and the BO rescore cache.
data/models/	166 MB, 115 files	Trained networks: the n = 36 samplers, the n = 180 policies, screens and learning rivals, and the n = 300 learning arms.
data/keys/landscape_l6m3_w3.csv	87 MB	The n = 36 answer key: one row per raw (A, B) pair of weight-3 polynomials over Z_6 x Z_3 (665,856 rows) with exact n, k, d, the merit kd^2/n and the canonical key.

In the n = 180 files, dX and dZ name each sector by its check matrix, so file dX is the paper's d_Z and file dZ the paper's d_X. Sign Z is gamma = +1 and sign X is gamma = -1.

The md5 of every file is in data/MANIFEST.md5 in the GitHub repository.

Versions

Each release of the repository reads one tag of this dataset. Tag v1.1 holds the data behind the full version of the paper, archived on Zenodo with DOI 10.5281/zenodo.23922222. Tag v1.0 holds the first release's data.

Usage

From a clone of the repository:

pip install -r requirements.txt
pip install huggingface_hub
python -m scripts.fetch_data      # downloads data/ at the release's tag and checks every md5
python -m pytest tests -q         # checks every readout against its expected output

The script downloads only data/, so this card and its license file never overwrite the repository's own. To download by hand, from the repository root:

hf download NunoBatista/gflowqec-data --repo-type dataset --revision v1.1 --include "data/*" --local-dir .

License

CC BY 4.0. Attribution: "GFlowQEC data, (c) 2026 Nuno Batista, licensed under CC BY 4.0." The code is MIT, licensed separately in the GitHub repository. To cite, use the repository's CITATION.cff.