Skip to content

Dataset · HuggingFaceH4

Filtered dialogues for SFT chat models

HuggingFaceH4/ultrachat_200k

200k filtered English dialogues used to train Zephyr-7B-β.

What's inside

This is a heavily filtered version of UltraChat: from 1.4 million ChatGPT-generated dialogues, a subset was selected for faster SFT. The set contains English dialogues on a wide range of topics; size ranges from 100k to 1M. The dataset was used to train Zephyr-7B-β. The format is text dialogues.

What you can do with it

  • Train a chat model or LoRA on SFT dialogues
  • Verify the quality of instruction fine-tuning
  • Compare with other SFT datasets for chatbots
Tasks
Text generation
Languages
English
Size
100K–1M rows
Downloads this month
87K
Likes
1.2K

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("HuggingFaceH4/ultrachat_200k")