README.md

metadata

language:
  - en
license: other
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
license_name: other
license_link: https://huggingface.co/datasets/stanfordnlp/coqa#licensing-information
dataset_info:
  features:
    - name: story
      dtype: string
    - name: previous_qas
      list: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: old_support_sentence
      dtype: string
    - name: new_support_sentences
      list: string
    - name: source
      dtype: string
    - name: open_domain_question
      dtype: string
  splits:
    - name: train
      num_bytes: 3068254
      num_examples: 1317
    - name: validation
      num_bytes: 1054840
      num_examples: 464
  download_size: 2649468
  dataset_size: 4123094
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*

[Under Review]

Update (01/05/2026): add open-domain question in the training set.

Update (11/19/2025): add open-domain question in the validation set.

This dataset is built upon the CoQA dataset (url: https://huggingface.co/datasets/stanfordnlp/coqa).

To avoid data contamination, however, the new support sentences generated by workers should be used solely for LLM inference unless a publicly available technical report or manuscript explicitly mentions they are used for training to ensure fairness in LLM evaluations.