UltraFeedback
OpenBMB (Tsinghua University)
64,000 prompts, four model answers each, scored by GPT-4 on four aspects.
Prompts were drawn from sources such as UltraChat, ShareGPT, Evol-Instruct, TruthfulQA, FalseQA, and FLAN, then answered by multiple LLMs. One row is a prompt with its four completions and fine-grained AI ratings with critiques. It is widely used to build preference pairs for reward models and DPO.
Access
- Cost
- FreeNo charge.
- License
- Open licenseMIT
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadParquet, JSON Lines
What's in it
- One row is
- One row per prompt with four rated completions
- Covers
- Not geographic
- Years
- 2023
- Updated
- One-time release
- Size
- About 64,000 prompts with 4 responses each
Quality
Influential for open preference tuning, but all ratings are by GPT-4 and known score-bias problems exist.
- Annotator is an LLM with length and style bias
- Some overall scores were later found inconsistent
- Terms of the generating models may apply
Connects to
- UltraChat: One source of prompts
- HelpSteer2: Human-annotated alternative for reward modeling
Questions it could answer
- How often does GPT-4 prefer the longest answer?
- How much do scores disagree with a human-rated set like HelpSteer2?