MedQA: Medical exam Q&A benchmark

Knowledge

A Q&A benchmark with questions collected from professional medical board exams. Only includes the English subset of the dataset (which also contains Mandarin Chinese and Taiwanese questions).

Contributed By

@bunny-baxter

@JasonBenn

Code

src/inspect_evals/medqa

Paper

https://arxiv.org/abs/2009.13081

Overview

MedQA is an evaluation dataset of medical knowledge questions collected from medical board exams. The full dataset has questions in English, Mandarin Chinese, and Taiwanese, but this evaluation only includes the English subset.

This evaluation is likely close to saturation. According to a result from 2023, GPT-4 was already scoring 86.1% at that time. GPT-4.1 does not currently score much better. A quick examination of its incorrect answers revealed some possibly incorrect or arguable questions, so it is not clear how much higher a realistic score might be.

Usage

Notice: Issue with datasets version

As of version 4.0.0 of datasets, dataset scripts are no longer supported by the package. This causes issues with the med_qa dataset and prevents use. Using this eval requires having datasets <4.0.0 (which is currently specified in pyproject.toml)

Usage

First, install the dependencies:

uv sync

Then, evaluate against one or more models with:

uv run inspect eval inspect_evals/medqa --model openai/gpt-4o

After running evaluations, you can view their logs using the inspect view command:

uv run inspect view

If you don’t want to specify the --model each time you run an evaluation, create a .env configuration file in your working directory that defines the INSPECT_EVAL_MODEL environment variable along with your API key. For example:

INSPECT_EVAL_MODEL=anthropic/claude-3-5-sonnet-20240620
ANTHROPIC_API_KEY=<anthropic-api-key>

Options

You can control a variety of options from the command line. For example:

uv run inspect eval inspect_evals/medqa --limit 10
uv run inspect eval inspect_evals/medqa --max-connections 10
uv run inspect eval inspect_evals/medqa --temperature 0.5

See uv run inspect eval --help for all available options.

Dataset

All questions are multiple choice, with 5 options and 1 correct answer. For humans, a passing score is 60%, and the 95th percentile score is 87%.

Example question:

  • Question: A scientist is studying the properties of myosin-actin interactions in a sample of human muscle tissue. She has identified a drug that selectively inhibits phosphate release by the myosin head. If she gives this drug to a sample of human muscle tissue under physiologic conditions, which of the following steps in cross-bridge cycling will most likely be blocked?
  • Choices:
      1. Myosin head release from actin
      1. Myosin head cocking
      1. Exposure of myosin-binding sites on actin
      1. Myosin head binding to actin
      1. Power stroke
  • Answer: E

Scoring

Simple accuracy over multiple choice answers.

Evaluation Report

Results from July 3, 2025, running on all 1,273 samples:

Model Accuracy
gpt-4o-mini 68.5%
gpt-4.1 87.4%