README.md

metadata

license: apache-2.0
task_categories:
  - question-answering
dataset_info:
  features:
    - name: problem_id
      dtype: string
    - name: modified_question
      dtype: string
    - name: original_question
      dtype: string
    - name: original_solution
      dtype: string
    - name: modified_solution
      dtype: string
    - name: original_answer
      dtype: string
    - name: modified_answer
      dtype: string
  splits:
    - name: train
      num_bytes: 69968
      num_examples: 80
  download_size: 18113
  dataset_size: 69968
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

ReasoningTrap: Diagnosing Instruction Overriding in Reasoning Models

This repository contains datasets designed to diagnose reasoning rigidity in large language models. These datasets are specially crafted modifications of existing mathematical benchmarks and puzzles to reveal how models override instructions. The paper describing this work can be found here: Reasoning Model is Stubborn: Diagnosing Instruction Overriding in Reasoning Models.

Project Page: https://ReasoningTrap.github.io

The datasets are available on the Hugging Face Hub:

• ConditionedMath (AIME): https://huggingface.co/datasets/ReasoningTrap/AIME
• ConditionedMath (MATH500): https://huggingface.co/datasets/ReasoningTrap/MATH500
• PuzzleTrivial: https://huggingface.co/datasets/ReasoningTrap/PuzzleTrivial