Skip to content

Dataset · argilla

Preference dataset based on Capybara dialogues

argilla/Capybara-Preferences

English multi-turn dialogues for LLM alignment and preference training (RLHF/DPO).

What's inside

Inside are dialogues in English collected based on the LDJnr/Capybara instructional dataset using the distilabel tool. The last assistant response was removed from the original conversations to turn the chat history up to the last user question into a base for preference evaluation. The set volume is between 10 and 100 thousand records in text and tabular form.

What you can do with it

  • Fine-tuning language models using RLHF or DPO on multi-turn dialogues
  • Training reward models to evaluate the quality of assistant responses
Tasks
Text generation
Languages
English
Size
10K–100K rows
Downloads this month
285
Likes
48

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("argilla/Capybara-Preferences")