Skip to content

Dataset · OpenMOSS-Team

MOSS-002 dialog SFT data

OpenMOSS-Team/moss-002-sft-data

An open dialog dataset for SFT: text-davinci-003 prompts and responses in English and Chinese.

What's inside

An open conversational dataset used to train MOSS-002. User prompts are expanded from a small set of human seed prompts using the Self-Instruct scheme, and AI responses are generated by text-davinci-003. The en_harmlessness prompts are taken from Anthropic red teaming data; the card mentions en_helpfulness, en_honesty, and other splits. Languages are English and Chinese.

What you can do with it

  • Dialog LLM SFT
  • LoRA fine-tuning for helpfulness/honesty
  • Safety evaluation on red-team prompts
Tasks
Text generation
Languages
English, Chinese
Size
1M–10M rows
Downloads this month
533
Likes
96

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("OpenMOSS-Team/moss-002-sft-data")