josephofthebread's picture
Release MedCalc-Bench-Verified-PII
ae2c9c5
|
Raw History Blame Contribute Delete
7.03 kB
metadata
license: cc-by-sa-4.0
task_categories:
  - text-generation
  - token-classification
language:
  - en
tags:
  - medical
  - privacy
  - memorization
  - personal-information
  - synthetic
  - clinical-notes
size_categories:
  - 10K<n<100K
pretty_name: MedCalc-Bench-Verified-PII
configs:
  - config_name: luna_terra
    data_files:
      - split: train
        path: luna_terra/train.parquet
      - split: test
        path: luna_terra/test.parquet
  - config_name: gpt_oss
    data_files:
      - split: train
        path: gpt_oss/train.parquet
      - split: test
        path: gpt_oss/test.parquet

MedCalc-Bench-Verified-PII

MedCalc-Bench-Verified with a made-up identity written into each patient note: the patient's name, date of birth, medical record number and home address. It is built to measure whether a model fine-tuned on the notes memorizes the patient's identity, and whether text sanitization removes it.

The dataset keeps MedCalc-Bench-Verified's train and test splits, rows, order and columns (revision 58fc3de). Only Patient Note changes: it holds the note with the identity written in. Each detail is added with the smallest possible edit, so the medical content, questions and answers are unchanged. Two columns are added: Pasted Details and Extraction Answers.

Configs

Config Name, date of birth, record number Address Extraction answers
luna_terra gpt-6-luna gpt-6.1-terra gpt-6-luna
gpt_oss gpt-oss-120b gpt-oss-120b gpt-oss-120b

Both configs cover the same 9,901 unique notes. The identity values are drawn per note with the same seed, so they mostly coincide: names differ in 127 notes and addresses in 277, where the two extraction models disagree; dates of birth and record numbers are identical. The configs differ in how the details were written into the text. 9,776 notes have all four details; 125 have none, because the patient is an animal (39) or the calculator values point to no single person (86). A note shared by several MedCalc rows has the same identity in each.

Columns

All MedCalc-Bench-Verified columns, plus:

Column Meaning
Pasted Details The edits that wrote the identity into Patient Note. Each edit has Details (name, address, dob, mrn), Values (the exact values), Start and End (its text in Patient Note), Original (the text it replaced, empty for an insertion) and Sources (the model that wrote it, or claude-sonnet-agent). An edit lists two details when one was written inside the other. Empty for notes without an identity
Extraction Answers JSON string with the answers to the 24 questions that shaped the identity (target person, sex, age, origin, places, tense, ...), with quotes from the note

Use

from datasets import load_dataset

ds = load_dataset("josephofthebread/MedCalc-Bench-Verified-PII", "gpt_oss", split="test")
row = ds[0]
print(row["Patient Note"][:400])

def remove(row, details):
    note = row["Patient Note"]
    for edit in sorted(row["Pasted Details"], key=lambda e: -e["Start"]):
        if set(edit["Details"]) & set(details):
            note = note[: edit["Start"]] + edit["Original"] + note[edit["End"] :]
    return note

original = remove(row, ["name", "address", "dob", "mrn"])
no_address = remove(row, ["address"])

Removing all details always gives back MedCalc's note exactly. Removing only some gives clean text when the edits of the removed details do not touch the kept ones; this holds for every subset in 9,137 notes of luna_terra and 8,279 of gpt_oss.

How it was built

  1. Extraction. A model answers 24 questions about each note: which person the calculator values belong to, and the patient's sex, age, origin, places of residence and care, and more.
  2. Identity. The name comes from Faker or Wikidata name lists, in a locale that fits the patient's origin or country. The date of birth fits the patient's age on a made-up presentation date. The record number has 6 to 10 random digits. The address is a real street, with a real house number or one inside the street's range, in the place where the note says the patient lives or was treated; otherwise in the United States.
  3. Pasting. One model call per detail, each on the note left by the previous call, with a placement and introducing words chosen by code. Automatic checks reject any edit that changes other text, attaches the detail to someone else, or breaks the sentence; a rejected detail is asked again up to five or seven times. The few details still missing were written by Claude Sonnet agents and passed the same checks.

Quality

One human annotator; reviewer agents read the samples first and flagged possible problems.

Extraction answers are judged by paste compatibility: right when details built from them do not contradict the note. On 60 notes × 24 questions, gpt-oss-120b was right on 1,428 of 1,440 (99.2%) and gpt-6-luna on 1,430 of 1,440 (99.3%).

Pasted details, on 40 notes per config (20 random, 20 hard), 160 details each. A detail is rejected when its date of birth contradicts the note or when it is a bracket placed right after another bracket.

Accepted 95% interval
luna_terra 154 of 160 (96.25%) 92.1% to 98.3%
gpt_oss 152 of 160 (95.0%) 90.4% to 97.4%

Most rejected dates come from notes whose year could not be fixed, so the birth date was computed from a default year and can conflict with clues such as COVID-19 testing.

License

CC BY-SA 4.0, as required by MedCalc-Bench-Verified.

Citation

@misc{medcalcbenchverifiedpii2026,
  title        = {{MedCalc-Bench-Verified-PII}: Synthetic Patient Identities Written into Clinical Notes},
  author       = {Scherbakov, Dmitry and Rakhimov, Damir},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/josephofthebread/MedCalc-Bench-Verified-PII}}
}

@misc{medcalcbenchverified,
  title        = {{MedCalc-Bench Verified}},
  author       = {Khandekar, Nikhil},
  howpublished = {\url{https://huggingface.co/datasets/nsk7153/MedCalc-Bench-Verified}},
  note         = {Revision 58fc3de568bfc53a68f32327e58f60dc6eaa17f8}
}

@inproceedings{khandekar2024medcalcbench,
  title     = {{MedCalc-Bench}: Evaluating Large Language Models for Medical Calculations},
  author    = {Khandekar, Nikhil and Jin, Qiao and Xiong, Guangzhi and Dunn, Soren and Applebaum, Serina S. and Anwar, Zain and Sarfo-Gyamfi, Maame and Safranek, Conrad W. and Anwar, Abid A. and Zhang, Andrew and Gilson, Aidan and Singer, Maxwell B. and Dave, Amisha and Taylor, Andrew and Zhang, Aidong and Chen, Qingyu and Lu, Zhiyong},
  booktitle = {Advances in Neural Information Processing Systems (Datasets and Benchmarks Track)},
  year      = {2024}
}