Safety

RealToxicityPrompts

RealToxicityPrompts measures how often a model continues a sentence fragment with text a toxicity classifier flags as toxic, using prompts drawn from both toxic and innocuous starting text.

2,006items

42subjects

unknownlicense

safetydomain

textmodality

item-level responses released

Saturation status: No

Full data on Hugging Face Back to the gallery

Paper and credit

2020

RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models (opens in a new tab)

Original source

Authors

Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi +1 moreShow all

01Samuel Gehman
02Suchin Gururangan
03Maarten Sap
04Yejin Choi
05Noah A. Smith

Citation

BibTeX citationShow

BibTeX

1@inproceedings{gehman2020realtoxicityprompts,
2  title={RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models},
3  author={Gehman, Samuel and Gururangan, Suchin and Sap, Maarten and Choi, Yejin and Smith, Noah A.},
4  booktitle={Findings of EMNLP},
5  year={2020},
6}

Response matrix

Responses come from Stanford CRFM's HELM Classic leaderboard (v0.4.0) at 5 generations per prompt, temperature 0.9; the original metrics of Gehman et al. are defined over 25 nucleus-sampled generations, so the two are not directly comparable.

Loading response matrix…

lowhighUnobserved

Scale: Score in [0, 1]