README.md

metadata

language:
  - en
pretty_name: DetailBench
tags:
  - text
license: apache-2.0
task_categories:
  - text-generation
  - translation
size_categories:
  - n<1K

DetailBench

This is the dataset for DetailBench, which answers the question: "How good are current LLMs at finding small errors, when they are not explicitly asked to do so?"

Dataset Structure

• article_title: Name of the Wikipedia article the data is from
• original_text: Original excerpt from the given Wikipedia article
• modified_text: Modified version of the original text with a single error (one changed number) introduced
• original_number: The original number from the text (used for the LLM grader as context)
• modified_number: The modified number from the text (used for the LLM grader as context)
• change_position: The position of the changed number in the text (used for the LLM grader as context)
• target_language: The language the LLM to evaluate should translate the modified_text into

Implementation

We recommend the reference implementation provided in openbench to run this benchmark. Simple use bench eval detailbench --model <model_name>

License

Apache 2.0