SKU PRF-3005 · Sold by External
HelpSteer2
Product specifications
| SKU | PRF-3005 |
|---|---|
| Data type | Preference data |
| Volume | 21,362 annotated prompt-response samples (20,324 train + 1,038 validation) |
| Size on disk | ~94 MB (HF usedStorage: 98,970,838 bytes across all files/splits) |
| Format | JSON Lines (gzipped .jsonl.gz; loadable via HF datasets) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open dataset (HF), CC-BY-4.0 |
| Quality score | — |
| License | Creative Commons Attribution 4.0 International (CC-BY-4.0) |
HelpSteer2 is an open-source helpfulness dataset released by NVIDIA in partnership with Scale AI to train reward models and align LLMs. It contains 21,362 samples (20,324 train + 1,038 validation), each pairing a prompt with a single response and five human-annotated attribute scores (helpfulness, correctness, coherence, complexity, verbosity), each ranging 0-4 where higher is better. Consecutive samples share the same prompt, so response pairs can be used for preference/DPO or Bradley-Terry reward modelling by comparing helpfulness scores. On 1 Oct 2024 NVIDIA additionally released a dedicated 'preference' split with human pairwise annotations (preference_strength from -3 to +3) plus written justifications, and a 'disagreements' split exposing individual per-annotator labels. When used to tune Llama-3.1-70B-Instruct, the resulting reward model reached 94.1% on RewardBench (best as of 1 Oct 2024, per the card). The dataset is permissively licensed CC-BY-4.0 and is not gated.