External dataset catalog
Other open conversation datasets. Being on Hugging Face doesn’t mean a dataset may be republished, so each one was reviewed for license, access and privacy terms. This is not legal advice.
| Dataset | License | Status | Reported size | Mirrored here |
|---|---|---|---|---|
| Anthropic HH-RLHF Not gated. Crowdworker conversations with Anthropic models. Red-team split contains harmful content; excluded from any seed. | MIT | Eligible to mirror | — | 247,333 |
| Arena Human Preference 55K Publisher-declared Apache-2.0 55K release. Both conversation sides retain exact source model IDs. Separate from the 100K/140K releases and their different terms. | Apache-2.0 | Eligible to mirror | 57,477 | 112,837 |
| Gata ChatGPT-RealUser-2.2M (preview) Not gated. 600-conversation public preview of real ChatGPT users who opted in via Gata's GPT-to-Earn program. The full 2.2M release has no public license and is not imported. | Apache-2.0 | Eligible to mirror | 600 | 590 |
| LMArena Human Preference 140K Not gated. Real Arena users' prompts (CC-BY-4.0) with both model responses. Publisher states model outputs are governed by each provider's terms; treated like the 55K release. Conversations with images are skipped. | Prompts CC-BY-4.0; outputs governed by provider terms | Eligible to mirror | 135,634 | 266,609 |
| LMArena Search Arena 24K Not gated. Real users' multi-turn chats with search-augmented models, both sides retained. Outputs governed by provider terms. | Prompts CC-BY-4.0; outputs governed by provider terms | Eligible to mirror | 24,069 | 47,287 |
| NVIDIA HelpSteer2 Human-rated evaluation responses. Source does not identify the individual NVIDIA model for each response; model identity is explicitly unspecified. Prompts are mostly ShareGPT-derived. | CC-BY-4.0 | Eligible to mirror | 21,362 | 21,264 |
| OpenAssistant Conversations (oasst2) Not gated. Assistant turns are human-written demonstrations by volunteers, not model outputs; labelled as such. | Apache-2.0 | Eligible to mirror | 13,854 | 79,740 |
| ShareLM (browser-extension contributions) Not gated. Only conversations users donated through the ShareLM browser extension, under its terms of use. Re-hosted WildChat, HH-RLHF, PRISM and other third-party subsets in ShareLM are skipped. Model is recorded as unspecified when only the chat site is known. | MIT | Eligible to mirror | — | 2,868 |
| WildChat-4.8M (AI2) Not gated. Public release is the non-toxic subset (3,199,860 of 4,804,190 original conversations). Upstream honours removal requests; The Chat Archive mirrors them via the suppression registry. | ODC-BY-1.0 | Eligible to mirror | 3,199,860 | 853,949 |
| Chatbot Arena Conversations Gated. Non-commercial terms on model outputs; at most a separate NC partition after review. Listed for discovery only. | Prompts CC-BY-4.0; outputs CC-BY-NC-4.0 | Catalog only | 33,000 | 0 |
| LMSYS-Chat-1M Gated. The license agreement forbids hosting or redistributing the dataset. Listed for discovery only. | LMSYS-Chat-1M Dataset License Agreement | Catalog only | 1,000,000 | 0 |
| ShareGPT (Vicuna unfiltered) Scraped from public share links without contributor consent; the uploader does not hold rights. Listed for discovery only. | Uploader-asserted Apache-2.0 (provenance unclear) | Catalog only | 53,000 | 0 |
| PRISM Alignment Dataset Not gated. 1,500 participants from 75 countries talking with 21 LLMs. Model responses are non-commercial; shipped in their own NC partition. The participant's chosen response is kept at each turn. | Human text CC-BY-4.0; model responses CC-BY-NC-4.0 | Separate license | 8,011 | 7,910 |