ToxicityPrompts/RealToxicityPrompts

Dataset Overview

Dataset Name

• Name: Real Toxicity Prompts

Basic Information

• Language: English
• License: Apache‑2.0
• Multilinguality: Monolingual
• Size: 100 K < n < 1 M
• Source: Original data
• Task Category: Text generation
• Labels: Toxicity, not suitable for all audiences

Dataset Description

• Summary: RealToxicityPrompts is a dataset of 100 k sentence fragments designed to help researchers further mitigate neural toxicity degeneration risks in language models.
• Language: English

Structure

• Data Instances: Each instance contains a prompt and its metadata, including filename, start/end offsets, challenge flag, prompt text and its toxicity scores (e.g., profanity, sexual insinuation, identity attack), as well as continuation text and its corresponding toxicity scores.

Creation

• Selection Rationale: Sentences were selected from the OPEN‑WEBTEXT CORPUS, with toxicity scores extracted via the Perspective API to obtain prompts across a spectrum of toxicity levels.
• License Information: Image metadata in the dataset is released under the Apache license.

Citation

@article{gehman2020realtoxicityprompts,
  title={Realtoxicityprompts: Evaluating neural toxic degeneration in language models},
  author={Gehman, Samuel and Gururangan, Suchin and Sap, Maarten and Choi, Yejin and Smith, Noah A},
  journal={arXiv preprint arXiv:2009.11462},
  year={2020}
}