Skip to content

Dataset · nvidia

Data for training reward models

nvidia/HelpSteer2

English helpfulness dataset for aligning LLMs on factuality, coherence, and response length.

What's inside

Open Helpfulness Dataset under the CC-BY-4.0 license; contains text and tabular data for model alignment. Enables fine-tuning for response helpfulness, factual accuracy, and coherence, as well as complexity and verbosity. Created in collaboration with Scale AI; volume is 10k-100k examples, language is English.

What you can do with it

  • Train a reward model for RLHF
  • Fine-tune Llama 3.1 70B Instruct
  • Evaluate on RewardBench
Languages
English
Size
10K–100K rows
Downloads this month
20K
Likes
468

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/HelpSteer2")