> ML_DATASET // ULTRAFEEDBACK-BINARIZED_v1.0
UltraFeedback: High-Quality Multi-Domain Preference Dataset
OpenBMB / Tsinghua University (Cui et al.) · RLHF & Preference Alignment · 64,000 prompts with 256,000 model completions evaluated by GPT-4
RLHF & Preference AlignmentMIT64,000 prompts with 256,000 model completions evaluated by GPT-4open
Dataset Profile & Characteristics
Label Type:Preference pairs (chosen vs rejected) with fine-grained multi-aspect numerical scores
Languages:en
License Tier:permissive-open-source
Modalities:text
Intended Use
- Training Direct Preference Optimization (DPO) and reward models (Zephyr-7B baseline)
Prohibited / Discouraged Use
- Evaluating model safety without additional red-teaming sets
Bias, Leakage & Privacy Risk Analysis
Privacy / Sensitive Data Risks:
Synthetic instructional prompts; zero personal data.
Known Bias:
GPT-4 judge bias (preference for longer, formatting-rich, polite responses).
Known Benchmark Leakage:
Prompt de-duplication against standard evaluation benchmarks.