Skip to content

README.md

Latest commit

 

History

History
42 lines (24 loc) · 2.6 KB

README.md

File metadata and controls

42 lines (24 loc) · 2.6 KB

Biomedical question answering robustness benchmark

The repo includes code to assess the robustness of large language models (LLMs) in biomedical question answering (QA). A collection of generalist and specialist LLMs are assessed.

This repo contains only the code for sampling-based attack, for running the benchmarks using the variant of discreteZOO, the zeroth-order optimization-based attack used in the manuscript, please see discretezoo-span.

Installation

Clone the code repo. Go to the qstab-main folder, install using the following from the command line

 pip install .

Run baseline (no attack)

The baseline assesses the zero-shot performance of LLMs for the specified datasets. Under the qstab-main folder, type the following from the command line

Run Flan-T5-large model on MedQA-USMLE drug-mention questions

 python ./scripts/run_hf_entity.py -grp="drugs" -nq=6000

Run MedAlpaca-7B model on MedQA-USMLE disease-mention questions

 python ./scripts/run_medalpaca_entity.py -grp="diseases" -nq=6000

Here 6000 is just a large enough number such that all instances are run, a smaller number evaluates only a subset of the data.

Run entity substitution attack

The substitution attack targets entities in the distractors of a question by constructing adversarial distractors. For each data instance, the attacker selects a replacement entity from a perturbations set using a sampling-based approach to substitute the original entity in the distractor.

  • Random sampling

An example to attack the MedAlpaca-7B model on MedQA-USMLE drug-mention questions using random sampling.

 python ./scripts/run_medalpaca_entity.py -ptb=True -ptb_samp="random" -grp="drugs" -nq=6000
  • Powerscaled distance-weighted sampling (PDWS)

An example to attack the MedAlpaca-7B model on MedQA-USMLE drug-mention questions using PDWS.

 python ./scripts/run_medalpaca_entity.py -ptb=True -ptb_samp="distance" -grp="drugs" -ndist=20 -nq=6000
  • Hyperparameter tuning for PDWS

An example to tune the PDWS attacker's hyperparameter for the MedAlpaca-7B model on MedQA-USMLE drug-mention questions

 bash ./scripts/medalpaca_tune.sh