Skip to content

Dataset · lmarena-ai

Chatbot Arena human preference battles

lmarena-ai/arena-human-preference-55k

55k real dialogues and preference evaluations between 70+ LLMs in pairwise battles.

What's inside

The dataset was collected for a Kaggle competition on predicting human preferences in Chatbot Arena. It features over 55,000 real user dialogues with LLMs and preference labels across 70+ models, including GPT-4, Claude 2, Llama 2, Gemini, and Mistral. Each example represents a pairwise battle of two LLMs on a single question with labels: choose A, B, tie, or both are bad. Data is tabular and text, in English.

What you can do with it

  • Train a preference classifier for LLMs
  • Build a reward model for RLHF
  • Evaluate chat models in pairwise comparisons
Tasks
Text classification
Languages
English
Size
10K–100K rows
Downloads this month
11.3K
Likes
170

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("lmarena-ai/arena-human-preference-55k")