license: cc-by-4.0
dataset_info:
features:

• name: id
  dtype: string
• name: type
  dtype: string
• name: prompt
  dtype: string
• name: completion
  dtype: string
• name: annotation_1
  dtype: string
• name: annotation_2
  dtype: string
• name: agreement
  dtype: bool
• name: final_label
  dtype: string
  splits:
• name: gpt4
  num_bytes: 264757
  num_examples: 450
• name: llama2new
  num_bytes: 523074
  num_examples: 450
• name: llama2orig
  num_bytes: 476580
  num_examples: 450
• name: mistralguard
  num_bytes: 309478
  num_examples: 450
• name: mistralinstruct
  num_bytes: 377642
  num_examples: 450
• name: prompts
  num_bytes: 47890
  num_examples: 450
  download_size: 930253
  dataset_size: 1999421
  configs:
• config_name: default
  data_files:
  • split: gpt4
    path: data/gpt4-*
  • split: llama2new
    path: data/llama2new-*
  • split: llama2orig
    path: data/llama2orig-*
  • split: mistralguard
    path: data/mistralguard-*
  • split: mistralinstruct
    path: data/mistralinstruct-*
  • split: prompts
    path: data/prompts-*
    task_categories:
• conversational
  language:
• en
  size_categories:
• 1K<n<10K

XSTest Dataset for Testing Exaggerated Safety

Note, this is an upload of the data found here for easier research use.
All credit to the authors of the paper

The test prompts are subject to Creative Commons Attribution 4.0 International license. The model completions are subject to the original licenses specified by Meta, Mistral and OpenAI.

Loading the dataset

Use the following:

from datasets import load_dataset
dataset = load_dataset("natolambert/xstest-v2-copy)
prompts = dataset["prompts"] # select the correct subset(s)

The dataset is a DatasetDict with the following structure:

DatasetDict({
    gpt4: Dataset({
        features: ['id', 'type', 'prompt', 'completion', 'annotation_1', 'annotation_2', 'agreement', 'final_label'],
        num_rows: 450
    })
    llama2new: Dataset({
        features: ['id', 'type', 'prompt', 'completion', 'annotation_1', 'annotation_2', 'agreement', 'final_label'],
        num_rows: 450
    })
    llama2orig: Dataset({
        features: ['id', 'type', 'prompt', 'completion', 'annotation_1', 'annotation_2', 'agreement', 'final_label'],
        num_rows: 450
    })
    mistralguard: Dataset({
        features: ['id', 'type', 'prompt', 'completion', 'annotation_1', 'annotation_2', 'agreement', 'final_label'],
        num_rows: 450
    })
    mistralinstruct: Dataset({
        features: ['id', 'type', 'prompt', 'completion', 'annotation_1', 'annotation_2', 'agreement', 'final_label'],
        num_rows: 450
    })

})

The mapping for prompts is as follows (from the original below)

• id_v1 removed,
• id_v2 -> id
• empty column completion
• focus to annotation_1 and note to annotation_2
• empty final_label

    prompts: Dataset({
        features: ['id_v1', 'id_v2', 'type', 'prompt', 'focus', 'note'],
        num_rows: 450
    })

Please cite their work:

@misc{röttger2023xstest,
      title={XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models}, 
      author={Paul Röttger and Hannah Rose Kirk and Bertie Vidgen and Giuseppe Attanasio and Federico Bianchi and Dirk Hovy},
      year={2023},
      eprint={2308.01263},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}