Spiderman01/MedPriv-Bench_dataset

MedPriv-Bench

MedPriv-Bench evaluates the privacy--utility trade-off of language models in medical open-ended question answering. Each example contains a synthetic patient context, injected privacy-sensitive facts, a question, and a ground-truth answer. The benchmark supports evaluating whether a model remains clinically useful while avoiding disclosure of the injected facts.

Split sizes

Split	Rows	Composition
train	2,200	1,315 benchmark-construction examples + 885 SFT examples
test	738	Held-out benchmark evaluation examples
Total	2,938	Matches the MedPriv-Bench paper

split_origin records the source subset of each row: construction_train_1315, sft_train_885, or heldout_test_738.

Important split note

This is a row-level split. Some patient_id values recur in the train and test partitions because distinct synthetic question/context instances were created from the same underlying source patient. The release contains 555 overlapping patient_id values across splits, but no identical full benchmark instances. It must therefore not be treated as a patient-level-disjoint generalization benchmark.

Fields

• record_id: stable release-local row identifier.
• patient_id, gender, age: source-patient metadata; patient_id is not a globally meaningful identifier.
• original_full_text, original_chunk_1--original_chunk_3: original context supplied to the model.
• injected_chunk_1--injected_chunk_3: context containing the synthetic sensitive facts used to test contextual leakage.
• phi_categories, phi_contents, phi_categories_json: annotations for the injected facts. These strings are the reference facts for leakage evaluation.
• user_question, question_reasoning_path, ground_truth_answer: task prompt, construction trace, and utility reference answer.
• verdict, verdict_qwen_max, verdict_gpt52: source construction/evaluation annotations when available; missing values mean the annotation was not part of that source subset.
• split_origin: original release subset, as above.

Intended use

Use the test split to evaluate an answer generator given the benchmark context and question. Score utility against ground_truth_answer and privacy against phi_contents / phi_categories_json with the accompanying MedPriv NLI judge. The dataset does not contain model answers; users generate those answers during evaluation.

Limitations and responsible use

The records are research data derived from clinical case material and contain synthetic sensitive attributes for evaluating privacy leakage. They are not for clinical decision-making, re-identification, patient profiling, or model deployment. Do not combine them with external data to attempt re-identification. The train/test split is row-level rather than patient-level, as described above.

License and attribution

This dataset is released under CC BY-NC-SA 4.0. It is a non-commercial research derivative of PMC-Patients, whose source project is also distributed under CC BY-NC-SA 4.0. Please preserve attribution and the same non-commercial, share-alike terms in derivatives.

Citation

@inproceedings{guan2026medprivbench,
  title={MedPriv-Bench: Benchmarking the Privacy-Utility Trade-off of Large Language Models in Medical Open-End Question Answering},
  author={Guan, Shaowei and Zhai, Yu and Kwok, Hin Chi and Du, Jiawei and Feng, Xinyu and Li, Jing and Qin, Harry and Hui, Vivian},
  year={2026}
}

Uploading this directory

The Parquet files and metadata are already prepared for Hugging Face. After creating an empty dataset repository, run:

python upload_to_hub.py --repo-id YOUR_ORG/MedPriv-Bench --private

Remove --private only when you are ready to publish. prepare_dataset.py rebuilds the Parquet files from the three canonical CSV inputs, and verify_dataset.py checks the release invariants.