WildChat-4.8M-Full
- Type
- dataset
- Venue
- Hugging Face / Allen Institute for AI (AI2)
- Year
- 2026
- Source
- huggingface
- Access
- restricted
- Language
- Multilingual (broad range of languages); primarily English
- Added
- 2026-07-17T20:18:03.644149+00:00
- Verified
- 2026-07-17T20:18:03.644149+00:00
Summary
WildChat-4.8M-Full is the complete (gated, Not-For-All-Audiences) version of a corpus of 4,743,336 real conversations between human users and ChatGPT (after removing minors' data from 4,804,190 original conversations), collected by offering free access to ChatGPT/GPT-4 in exchange for consensual chat-history sharing. It includes both toxic and non-toxic conversations (the non-toxic-only version is WildChat-4.8M with 3,199,860 conversations) and contains interactions with reasoning models o1-preview and o1-mini. Each conversation includes role, content, detected language, toxicity flags, hashed IP, state/country, and request headers.
Keywords
conversations chatgpt instruction-tuning real-world multilingual allenai gated sensitive
Topics
NLP / Conversational AI
Research notes
- Marked Not-For-All-Audiences on HuggingFace due to sensitive/toxic content and PII. The non-toxic filtered version is allenai/WildChat-4.8M. Papers: arxiv 2405.01470, 2409.03753, 2406.04770.