BAND-MedQA: Demographically Perturbed Medical Question Answering

This dataset accompanies BAND: Bias-Aware Retrieval Debiasing for Demographic Robustness in Retrieval-Augmented Medical Question Answering, accepted at the GroundLM Workshop @ EMNLP 2026.

BAND-MedQA is an ethnicity-perturbed version of MedQA designed to study retrieval shift: changes in biomedical evidence retrieval caused by demographic information that does not alter the clinically correct answer. It supports analysis of demographic robustness in retrieval-augmented medical question answering (RAG).

Code: BAND GitHub repository

Dataset Construction

Starting from MedQA, we retain questions for which an ethnicity perturbation should not change the correct answer. The paper describes two filtering stages:

1. Keyword filtering: Exclude questions containing explicit ethnicity terms.
2. LLM-based screening: Use GPT-4o-mini to exclude questions where changing ethnicity could change the gold answer.

After filtering, 11,179 questions (87.9%) remain across the original train, development, and test splits. For each retained question, a standardized sentence is prepended to the clinical question for one of five ethnicity groups:

Ethnicity group	Example prefix
African American	The patient is of African descent.
Caucasian	The patient is of Caucasian descent.
Asian	The patient is of Asian descent.
Hispanic	The patient is of Hispanic descent.
Native American	The patient is of Native American descent.

The answer choices and gold answer are intended to remain unchanged across perturbations.

Dataset Statistics

Each retained MedQA question is distributed as six records: one original and five ethnicity-perturbed variants. All three uploaded JSONL files were validated, with no malformed JSON lines.

Split	Original questions	Perturbed questions	Total JSONL rows
Train	8,944	44,720	53,664
Development	1,120	5,600	6,720
Test	1,115	5,575	6,690
Total	11,179	55,895	67,074

The paper reports perturbed-question counts (44,720 / 5,600 / 5,575), whereas the released files also include original questions. Every original question has exactly five perturbations in its split, grouped by _group_id.

Files and Splits

• phrases_rethnicity_train_prefix_perturbed.jsonl — training data
• phrases_rethnicity_dev_prefix_perturbed.jsonl — development/validation data
• phrases_rethnicity_test_prefix_perturbed.jsonl — test data

To make the Hugging Face Dataset Viewer recognize the split names reliably, upload the files under data/train.jsonl, data/validation.jsonl, and data/test.jsonl, respectively, or configure explicit configs metadata in this card.

Data Format

All three files use JSON Lines (.jsonl) format, with one JSON object per line and the same 11 fields:

Field	Type	Description
question	string	Original question or question with a prepended ethnicity sentence
answer	string	Gold answer text
options	object	Multiple-choice answers keyed by option label (e.g., A–D)
meta_info	string	MedQA metadata (observed: step1, step2&3)
answer_idx	string	Correct option label
metamap_phrases	list of strings	Medical/concept phrases supplied with the question
is_perturbed	boolean	Whether an ethnicity prefix was added
original_ethnicity	null	Original ethnicity metadata; null in all three uploaded splits
swapped_ethnicity	string or null	Perturbation group identifier; null for originals
_group_id	integer	Connects each original question to its five perturbations within the file
final	boolean	Observed to match is_perturbed across all three splits; its intended independent meaning is not documented.

For example, the following shows a schematic excerpt (not a complete row):

{
  "question": "The patient is of African descent. [Original clinical question]",
  "answer_idx": "B",
  "is_perturbed": true,
  "original_ethnicity": null,
  "swapped_ethnicity": "african_american",
  "_group_id": 0
}

The train, development, and test files contain 8,944, 1,120, and 1,115 distinct _group_id values, respectively. Each appears exactly six times: once for the original and once for each ethnicity perturbation. Group identifiers are local to a split.

Loading the Dataset

Once the three files are uploaded to the Hugging Face dataset repository under data/train.jsonl, data/validation.jsonl, and data/test.jsonl, load them using:

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/BAND-MedQA")
print(dataset)

train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

original_test = test.filter(lambda row: not row["is_perturbed"])
perturbed_test = test.filter(lambda row: row["is_perturbed"])
print(len(original_test), len(perturbed_test))  # 1115 5575

To load the original filenames locally without renaming them:

from datasets import load_dataset

dataset = load_dataset("json", data_files={
    "train": "phrases_rethnicity_train_prefix_perturbed.jsonl",
    "validation": "phrases_rethnicity_dev_prefix_perturbed.jsonl",
    "test": "phrases_rethnicity_test_prefix_perturbed.jsonl",
})

Use is_perturbed and swapped_ethnicity to select demographic variants, and _group_id to match them to their original questions within each split.

Recommended Evaluation Protocol

The paper evaluates medical RAG systems on ethnicity-perturbed questions using:

• Accuracy: Correct-answer rate, including accuracy by ethnicity group.
• DP Gap: Difference between the highest and lowest group-level accuracy.
• Flip Rate: Proportion of originally correct predictions that become incorrect after ethnicity perturbation.
• Retrieval Consistency: Mean Jaccard similarity between evidence sets retrieved for original and perturbed questions.
• BiasedEvidence: Proportion of perturbed queries whose retrieved evidence contains any demographic terms in the predefined lexicon.
• AttrRate: Proportion of perturbed queries whose retrieved evidence contains terms matching the injected ethnicity group.
• DiffDocs: Proportion of perturbed queries whose retrieved document set differs from that of the original query.

Caution: The paper's BiasedEvidence and AttrRate are based on deterministic keyword matching against a predefined demographic lexicon; they are not human-annotated judgments of clinical irrelevance or fairness.

Intended Use and Limitations

This dataset is intended for research into demographic robustness, retrieval instability, and bias propagation in medical RAG systems. It is not intended for direct clinical decision-making.

The study evaluates ethnicity perturbations and does not establish robustness across other demographic attributes. The perturbations are designed to preserve the correct answer, but the resulting benchmark does not represent all situations in which demographic information is clinically relevant. Findings should not be interpreted as guarantees of fairness or safety in clinical deployment.

Source, Redistribution, and License

The dataset is derived from MedQA (Jin et al., 2021). Before publicly distributing the underlying questions and answer choices, verify the original dataset's redistribution terms and any applicable restrictions. No license is asserted in this card pending that review.

Citation

If you use this dataset, please cite the BAND paper:

Harim Lee, Dasom Lee, and Chaehyun Ryu. BAND: Bias-Aware Retrieval Debiasing for Demographic Robustness in Retrieval-Augmented Medical Question Answering. GroundLM Workshop @ EMNLP 2026.

Please also cite the original MedQA dataset paper (Jin et al., 2021). Add the official BAND BibTeX entry when publication metadata is available.

Acknowledgements

The BAND implementation builds on the rationale-guided medical RAG pipeline RAG² (Sohn et al., 2025). Dataset construction is based on MedQA (Jin et al., 2021).