M-ARC

HuggingFace upload of a clinical QA benchmark designed to exploit LLMs' "inductive biases toward inflexible pattern matching from their training data rather than engaging in flexible reasoning." If used, please cite the original authors using the citation below.

Dataset Details

Dataset Description

The dataset contains one split:

• test: up to seven-option multiple-choice QA (choices A-G)

Dataset Sources

• Repository: https://github.com/dbernardo05/medARC-QA
• Paper: https://arxiv.org/pdf/2502.04381

Direct Use

import json
from datasets import load_dataset

if __name__ == "__main__":
    # load the test split
    dataset_test = load_dataset("mkieffer/M-ARC", split="test")
    print("\test split:\n", dataset_test)
    print("\ntest sample:\n", json.dumps(dataset_test[0], indent=2))

Citation

@misc{kim2025limitationslargelanguagemodels,
      title={Limitations of Large Language Models in Clinical Problem-Solving Arising from Inflexible Reasoning}, 
      author={Jonathan Kim and Anna Podlasek and Kie Shidara and Feng Liu and Ahmed Alaa and Danilo Bernardo},
      year={2025},
      eprint={2502.04381},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.04381}, 
}