Dataset guide
Choose a WildChat release
Match the collection and filtering to the research question.
Start with the open WildChat-4.8M release for most new projects. Use a Full release when the study requires flagged content and access has been approved. For replication, match the repository, revision, and filters used by the original study.
Current downloads and access requests: yuntian-deng/WildChat-4.8M and yuntian-deng/WildChat-4.8M-Full. Request Full access in the latter repository.
Counts and access information checked October 19, 2026 against the linked dataset cards.
Available releases
The “4.8M” name identifies the release family rather than every download’s row count. “Full” still includes documented removals.
Open download; toxicity-filtered
The expanded human–ChatGPT collection, excluding conversations flagged by OpenAI Moderation or Detoxify.
- Current count
- 3,199,860 conversations in the current repository metadata. The 4.8M name identifies the release family.
- Collection coverage
- The 4.8M collection extends through July 31, 2025; the Full release card documents this collection cutoff.
- Models
- GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4o, GPT-4.1 mini, o1-mini, o1-preview
- Use when
- A starting point for most new projects that do not need flagged content.
Dataset card and count source
Gated; reviewed justification required
Includes flagged and unflagged conversations, with the stated category removal still applied. The card reports 122,921 o1-mini/o1-preview conversations.
- Current count
- 4,743,336 after removing conversations flagged in the sexual/minors category, from an original 4,804,190.
- Collection coverage
- Through July 31, 2025; updated August 11, 2025.
- Models
- GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4o, GPT-4.1 mini, o1-mini, o1-preview
- Use when
- Studies whose question requires flagged content, subject to the dataset access process.
Dataset card and count source
Working with the data
Choose and record a snapshot
Use the yuntian-deng repositories linked here for current downloads and access requests. For comparisons with earlier studies, match their collection window, revision, and filters. Release names are not current row counts. Source.
Request Full access in the current repository
Submit an access request on yuntian-deng/WildChat-4.8M-Full with a detailed explanation of why your work needs flagged content. Approval is reviewed manually. Source.
Keep evaluation data separate
The processing pipeline removes conversations reserved for WildBench. Preserve that separation when constructing training sets or comparing with the benchmark. Source.
Distinguish logs from synthetic extensions
WildChat-50M is a separate follow-up by Benjamin Feuer and Chinmay Hegde: it adds responses from more than 50 open-weight models to WildChat prompts. Its name does not mean 50 million newly collected human conversations. Source.
Use the appropriate citation
Cite Zhao, Ren, Hessel, Cardie, Choi, and Deng’s ICLR 2024 WildChat paper for the dataset. Cite the separate WildVis paper when using the visualizer, and identify the downloaded release and revision. Source.
Check terms and removal instructions
The Full release card lists ODC-BY and explains how to request removal using conversation hashes or turn identifiers and the contact details on Yuntian Deng’s homepage. Source.
Interpret metadata carefully
The paper describes geographic data inferred from IP addresses and discusses selection bias in the service’s user population. Treat this as a particular population of chatbot users when interpreting aggregate behavior. Source.
Understand the identifiers
In the expanded releases, conversation_hash identifies content and is not a unique conversation key. The cards identify turn_identifier as the unique identifier within each turn. Source.
Read the ODC-BY license and the Full release card for terms and collection details. The current removal-request instructions remain with the dataset.
Selected collection and release milestones
The original paper describes an approximately one-million-conversation corpus, and some experiments use earlier collection windows. The expanded 4.8M releases cover a longer period and more models. Historical study counts should be read in that context.
Collection dates, paper versions, and dataset updates describe different events. The arXiv version below does not establish the first public release of WildChat.
- April 9, 2023
Start of the collection window reported for the expanded corpus in the May 2024 paper. Source.
Collection date, not a public release date.
- May 2, 2024
The WildChat arXiv version describes the expanded one-million-conversation corpus. Source.
arXiv version date, distinct from the earlier ICLR submission and initial dataset release.
- August 11, 2025
The 4.8M release update extends coverage to before August 1, 2025 and adds secret scanning. Source.
Release-update date; the collection cutoff is earlier.
Download this release guide as JSON