HelpSteer2
NVIDIA (with Scale AI)
Prompt-response pairs rated on five attributes, small enough to inspect.
Human ratings of model replies for helpfulness, correctness, coherence, complexity, and verbosity, collected with Scale AI. One row is a prompt, a response, and the attribute scores. NVIDIA reports that a model trained on it reached top reward-model results at release.
Access
- Cost
- FreeNo charge.
- License
- Open licenseCC BY 4.0
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadJSON Lines, Parquet
What's in it
- One row is
- One row per prompt and response with attribute scores
- Covers
- Not geographic
- Years
- 2024
- Updated
- One-time release
- Size
- 10,000 to 100,000 rows
Quality
Professionally annotated with attribute-level scores, which is more useful than a single preference bit.
- Responses are model-generated
- Small enough that reward models can overfit
Connects to
- HH-RLHF (Helpful and Harmless): Earlier preference data with a binary choice
- UltraFeedback: Larger preference set with GPT-4 annotations
Questions it could answer
- Which attribute best predicts overall helpfulness?
- How correlated are verbosity and correctness scores?