What’s Out There
← Back to swiping
Language & textFreeAcademic

Chatbot Arena Conversations

LMSYS (UC Berkeley and collaborators)

33,000 side-by-side model chats where real users voted for the better answer.

Cleaned conversations from Chatbot Arena, April to June 2023, each with two anonymous models and the user's vote. One row is one head-to-head conversation with both model names, the full text, and the vote. It is the raw material behind Elo-style LLM rankings.

Access

Cost
FreeFree but gated; accept the terms on the Hub.
License
Provider's own termsTagged generic 'cc' in the record; the specific license needs checking
Sign-up
Free account
How you get it
API and Bulk downloadParquet

What's in it

One row is
One row per pairwise comparison
Covers
Global users, 13,000 unique IP addresses
Years
April to June 2023
Updated
One-time release
Size
33,000 conversations
model Amodel Bconversation textuser votelanguagemoderation tag

Quality

3 of 5

Real human votes, but a one-time 2023 snapshot with only 33,000 rows.

  • Voters are self-selected and prompts are uncontrolled
  • Votes favor style and length as well as accuracy
  • Models are all from 2023

Connects to

Questions it could answer

  • How often do ties and both-bad votes occur by model pair?
  • Does voter preference track answer length?

Similar data sets

  • LMSYS-Chat-1M

    One million real conversations with 25 LLMs, collected from Vicuna demo and Chatbot Arena visitors.

    Language & textLMSYS (UC Berkeley and collaborators)Free

  • Arena Leaderboard Dataset

    Historical snapshots of the Arena leaderboard, showing how model rankings moved over time.

    Internet & techLMArenaFree

  • WildChat-1M

    A million real ChatGPT conversations with state, country, and hashed-IP fields attached.

    Language & textAllen Institute for AI and University of WashingtonFree

  • OpenAssistant Conversations (OASST1)

    161,443 chat messages written and ranked by volunteers in 35 languages, with no model in the loop.

    Language & textOpenAssistant (LAION-affiliated volunteers)Free

  • UltraFeedback

    64,000 prompts, four model answers each, scored by GPT-4 on four aspects.

    Language & textOpenBMB (Tsinghua University)Free

  • UltraChat 200k

    A 200,000-dialogue filtered slice of ChatGPT-generated conversations that trained Zephyr-7B.

    Language & textHugging Face H4 team (filtered from Tsinghua's UltraChat)Free