Skip to main content
Datasetinstruction-tuningv1.0

UltraFeedback

by Tsinghua University · open-source · Last verified 2026-03-17

A large-scale, high-quality preference dataset with 64,000 instructions each answered by 4 LLMs and rated by GPT-4 on instruction-following, truthfulness, honesty, and helpfulness. UltraFeedback is the backbone of the Zephyr and Tulu 2 DPO models.

https://huggingface.co/datasets/openbmb/UltraFeedback
B+
B+Good
Adoption: B+Quality: AFreshness: B+Citations: AEngagement: F

Specifications

License
MIT
Pricing
open-source
Capabilities
reward-model-training, rlhf, dpo-training, preference-learning
Integrations
huggingface-datasets
Use Cases
rlhf, dpo, reward-modeling, alignment-research
API Available
No
Tags
rlhf, preference-data, gpt-4-annotated, reward-model, alignment
Added
2026-03-17
Completeness
100%

Index Score

70.2
Adoption
79
Quality
88
Freshness
74
Citations
84
Engagement
0

Put AI to work for your business

Deploy this dataset alongside autonomous AaaS agents that handle tasks end-to-end — no babysitting required.

Explore the full AI ecosystem on Agents as a Service