Skip to content

Dataset · LDJnr

Multi-turn GPT-4 human dialogues

LDJnr/Puffin

3,000 examples of GPT-4 responses across over 2,000 multi-turn English dialogues for chat model training.

What's inside

Official Puffin dataset: 3,000 examples with responses generated by GPT-4. Based on over 2,000 multi-turn conversations between GPT-4 and real humans; average context length exceeds 1,000 tokens, and average turns exceed 10. The author notes that Puffin is outdated and recommends Pure-Dove.

What you can do with it

  • Fine-tune a conversational LLM in SFT format
  • Evaluate models on long multi-turn chats
  • Gather LoRA for the GPT-4 response style
Tasks
Question answering, Text generation
Languages
English
Size
1K–10K rows
Downloads this month
542
Likes
97

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("LDJnr/Puffin")