Datasets:

---

ncbi

/

Open-Patients

like 25

Follow

NLM/DIR BioNLP Group 92

I was trying to map all the USMLE IDs back to their source MedQA questions and found that MedQA-USMLE contains a few duplicate pairs: rows that contain the same question and answer choice options, but with a different actual correct answer listed.

By running the sql query below in the open-patients HF data viewer

SELECT _id, description
FROM train
WHERE _id IN ('usmle-1100', 'usmle-1409', 'usmle-2038', 'usmle-2044')
LIMIT 100;

you will see two duplicate pairs that carried over into Open-Patients:

1. ['usmle-1100', 'usmle-1409']
2. ['usmle-2038', 'usmle-2044']

Here is the python code I used to find them:

from __future__ import annotations

import sys
from collections import defaultdict
from typing import Dict, List
from datasets import load_dataset


def main() -> int:
    ds = load_dataset("ncbi/Open-Patients", split="train")

    desc_to_ids: Dict[str, List[str]] = defaultdict(list)

    for row in ds:
        desc = row.get("description")
        rec_id = row.get("_id")
        if not isinstance(desc, str) or rec_id is None:
            continue
        desc_to_ids[desc].append(str(rec_id))

    dup_count = 0
    for desc, ids in desc_to_ids.items():
        if len(ids) >= 2:
            dup_count += 1
            print(
                f"duplicate_count={len(ids)} ids={ids}"
            )

    print(f"\nTotal duplicate description groups: {dup_count}")
    return 0

if __name__ == "__main__":
    raise SystemExit(main())