Toxic Chat
Toxic conversations collected from real interactions with a chat assistant.
Made by LMSYS OrgDatasets and benchmarks
About Toxic Chat
The dataset page is linked to the Large Model Systems Organization on Hugging Face. The dataset is listed under the task Text Classification.
The dataset page lists an associated arXiv paper, arxiv: 2310.17389.
Drafted from huggingface.co · not yet confirmed by LMSYS Org
How it works
ToxicChat is about 10,000 real user prompts to an open chatbot (Vicuna), annotated for toxicity and for jailbreaking attempts. Unlike social media datasets it captures what people say to an assistant, which is a different distribution: more requests, more attempts to trick the model, and more subtle toxicity.
It's the right evaluation set for a chatbot's input filter.
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing Toxic Chat? Send them this page.