Dataset, tools, and research
WildChat
Real conversations for studying how people use language models.
WildChat collects multilingual, multi-turn conversations between consenting users and ChatGPT. It supports research on user behavior, safety, evaluation, and instruction tuning.
Original paperChoose a dataset releaseResearch using WildChatExplore with WildVis
The original contribution · ICLR 2024
WildChat: 1M ChatGPT Interaction Logs in the Wild
Wenting Zhao and Yuntian Deng contributed equally.
Introduces consented, multilingual, multi-turn conversations with GPT-3.5 and GPT-4, together with timestamps and metadata for studying use across time and locations. The paper examines language and prompt diversity, toxicity and jailbreaks, and instruction tuning with WildLlama.
The collection pairs conversation text with timestamps, geographic metadata, hashed IP addresses, and request headers. These enable analyses of language-model use across conversations, locations, and time. Read the collection methods and analyses.
A guide to using WildChat
Dataset releases
The 4.8M releases extend the collection through July 2025. The open release filters flagged conversations; the Full release includes them and requires an access request.
Research using WildChat
Read summaries of work on privacy, safety, evaluation, information access, and synthetic training data, with dataset details and citations.
Explore conversations with WildVis
Search and visualize WildChat and LMSYS-Chat-1M with WildVis, using conversation filters and embedding maps.
Get started
Download WildChat-4.8M or compare releases. The processing repository documents conversation reconstruction, redaction, moderation, and reserved evaluation data.
Download the research annotations and citations as JSON, Markdown, BibTeX, or CSL-JSON.
Release information checked October 19, 2026. About this guide.