BAND-MedCPT

Pretrained debiased retrieval encoder for BAND: Bias-Aware Retrieval Debiasing for Demographic Robustness in Retrieval-Augmented Medical Question Answering.

Accepted at the GroundLM Workshop @ EMNLP 2026.

GitHub Repository | Dataset

Model Description

BAND (Bias-Aware Neutral Debiasing) is a retrieval debiasing framework designed to mitigate demographic-induced retrieval shift in retrieval-augmented medical question answering.

The framework addresses cases where demographic perturbations influence retrieved evidence despite the underlying clinical question remaining unchanged.

This repository provides the pretrained debiased retrieval encoder used in BAND.

Intended Use

The model is intended for research on:

• Demographic robustness in medical retrieval
• Retrieval-augmented medical question answering
• Retrieval representation debiasing
• Robustness evaluation under demographic perturbations

This model is intended for research use and is not validated for clinical decision-making.

Model Usage

Download the model files:

from huggingface_hub import snapshot_download

model_path = snapshot_download(
    repo_id="hrimlee/BAND-MedCPT",
    local_dir="outputs/medcpt_debias_merged"
)

print(model_path)

For inference with the BAND framework, refer to the official GitHub repository.

Evaluation

BAND is evaluated on ethnicity-perturbed MedQA, considering both answer-level robustness and retrieval-level behavior.

The evaluation includes:

• Answer accuracy
• Demographic parity gap
• Prediction flip rate
• Retrieval consistency
• Demographic keyword-based evidence metrics

For full experimental results, see the paper and official implementation.

Limitations

• The model was evaluated in a research setting using ethnicity-perturbed medical questions.
• Robustness under the evaluated perturbations does not guarantee fairness across other demographic attributes or clinical settings.
• The model should not be used as a standalone clinical decision-support system.

Citation

If you use this model, please cite:

BAND: Bias-Aware Retrieval Debiasing for Demographic Robustness in Retrieval-Augmented Medical Question Answering.

GroundLM Workshop @ EMNLP 2026.

Acknowledgements

This work builds upon RAG² (Sohn et al., 2025), Rationale-guided Retrieval Augmented Generation for Medical Question Answering.