README.md

metadata

license: mit
task_categories:
  - question-answering

This dataset accompanies the paper:

When Do LLMs Admit Their Mistakes? Understanding the Role of Model Belief in Retraction

It includes the original Wikidata questions used in our experiments, with train/test split. For a detailed explanation of the dataset construction and usage, please refer to the paper.

This repository provides:

• Original datasets: wikidata_{train,test}_free.jsonl, celebrity_{train,test}_free.jsonl.
• Continuation datasets: for Llama3.1-8B, Qwen2.5-7B, and Olmo2-7B.

Code: https://github.com/YuqingYang125/retraction

Citation

@misc{yang2025llmsadmitmistakesunderstanding,
      title={When Do LLMs Admit Their Mistakes? Understanding the Role of Model Belief in Retraction}, 
      author={Yuqing Yang and Robin Jia},
      year={2025},
      eprint={2505.16170},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.16170}, 
}