v1.1: inclusion criteria, leakage-aware splits, rebuild script for unlicensed sources
3fcd003 verified | """Rebuild the benchmark domains whose source corpora state no licence. | |
| Downloads each source from the authors' GitHub repository, matches every item of our | |
| split by the SHA-1 of its normalised text (manifests/), and writes | |
| data/<domain>/{train,test}.csv with our labels in our row order. | |
| Usage (from the release folder): python rebuild.py | |
| Requires: pandas, openpyxl | |
| """ | |
| import hashlib | |
| import io | |
| import tempfile | |
| import urllib.request | |
| import zipfile | |
| from pathlib import Path | |
| import pandas as pd | |
| from release_sources import DOMAIN_SOURCE, READERS, REPOS, norm | |
| HERE = Path(__file__).resolve().parent | |
| def fetch(repo, dest): | |
| url = f"https://codeload.github.com/{repo}/zip/HEAD" | |
| with urllib.request.urlopen(url) as r: | |
| zipfile.ZipFile(io.BytesIO(r.read())).extractall(dest) | |
| def main(): | |
| with tempfile.TemporaryDirectory() as tmp: | |
| texts = {} | |
| for src, repo in REPOS.items(): | |
| print(f"downloading {repo}") | |
| root = Path(tmp) / src | |
| fetch(repo, root) | |
| texts[src] = {hashlib.sha1(norm(t).encode("utf-8")).hexdigest(): t for t in READERS[src](str(root))} | |
| for key, src in DOMAIN_SOURCE.items(): | |
| for split in ("train", "test"): | |
| man = pd.read_csv(HERE / "manifests" / key / f"{split}.csv", dtype=str) | |
| man["text"] = man.sha1.map(texts[src]) | |
| missing = man.text.isna().sum() | |
| out = HERE / "data" / key / f"{split}.csv" | |
| out.parent.mkdir(parents=True, exist_ok=True) | |
| man.dropna(subset=["text"])[["text", "label"]].to_csv(out, index=False, encoding="utf-8") | |
| print(f" {key}/{split}: {len(man) - missing:,} of {len(man):,} items" | |
| + (f" ({missing} not found in the current source file)" if missing else "")) | |
| if __name__ == "__main__": | |
| main() | |