Datasets:
Tasks:
Text Retrieval
Formats:
parquet
Sub-tasks:
document-retrieval
Languages:
English
Size:
10M - 100M
License:
[Nikita's AI] Add examples, MANIFEST.json and SHA256SUMS for the v0.1-preview files
12e6aff verified | """Attach the original MS MARCO text to Doubting MS MARCO pairs, from the official TSV files. | |
| The package stores ids only. Download the MS MARCO passage ranking files yourself (https://microsoft.github.io/msmarco/): | |
| collection.tsv (passage id, text) and the queries file of the split (queries.train.tsv or queries.dev.small.tsv; | |
| query id, text). Only the needed lines are kept in memory. | |
| Usage: python join_text.py --package DIR --config qrels --split dev --collection collection.tsv | |
| --queries queries.dev.small.tsv --out pairs_with_text.tsv [--limit N] | |
| Writes query_id, passage_id, grade (and origin, agent or official, for --config judgments, where one pair can carry | |
| both), query text and passage text as TSV. Needs pyarrow (`pip install pyarrow`). | |
| """ | |
| import argparse | |
| import csv | |
| import glob | |
| import os | |
| import sys | |
| import pyarrow.parquet as pq | |
| def columns(config): | |
| return ["query_id", "passage_id", "grade"] + (["origin"] if config == "judgments" else []) | |
| def read_pairs(package, config, split, limit=None): | |
| """Rows of (query_id, passage_id, grade[, origin]) in file order.""" | |
| pairs = [] | |
| for path in sorted(glob.glob(os.path.join(package, "data", config, f"{split}-*.parquet"))): | |
| t = pq.read_table(path, columns=columns(config)) | |
| for row in zip(*(t[c].to_pylist() for c in columns(config))): | |
| pairs.append(row) | |
| if limit and len(pairs) >= limit: | |
| return pairs | |
| if not pairs: | |
| sys.exit(f"no Parquet files for {config}/{split} under {package}") | |
| return pairs | |
| def read_tsv_texts(path, wanted): | |
| """{id: text} for the wanted ids of a two-column TSV (id, text).""" | |
| out = {} | |
| with open(path, encoding="utf-8", newline="") as f: | |
| for line in f: | |
| i, _, text = line.rstrip("\n").partition("\t") | |
| if i in wanted: | |
| out[i] = text | |
| return out | |
| def join(pairs, collection, queries): | |
| p_text = read_tsv_texts(collection, {r[1] for r in pairs}) | |
| q_text = read_tsv_texts(queries, {r[0] for r in pairs}) | |
| missing_p = {r[1] for r in pairs} - set(p_text) | |
| missing_q = {r[0] for r in pairs} - set(q_text) | |
| if missing_p or missing_q: | |
| sys.exit(f"{len(missing_q)} query ids and {len(missing_p)} passage ids not found: wrong file or split?") | |
| return [tuple(r) + (q_text[r[0]], p_text[r[1]]) for r in pairs] | |
| def main(): | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--package", default=os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) | |
| ap.add_argument("--config", default="qrels", choices=["qrels", "judgments"]) | |
| ap.add_argument("--split", default="dev", choices=["train", "dev"]) | |
| ap.add_argument("--collection", required=True) | |
| ap.add_argument("--queries", required=True) | |
| ap.add_argument("--out", required=True) | |
| ap.add_argument("--limit", type=int) | |
| a = ap.parse_args() | |
| rows = join(read_pairs(a.package, a.config, a.split, a.limit), a.collection, a.queries) | |
| with open(a.out, "w", encoding="utf-8", newline="") as f: | |
| w = csv.writer(f, delimiter="\t", lineterminator="\n") | |
| w.writerow(columns(a.config) + ["query", "passage"]) | |
| w.writerows(rows) | |
| print(f"wrote {len(rows):,} rows to {a.out}") | |
| if __name__ == "__main__": | |
| main() | |