Skip to content

Dataset · stanfordnlp

385K human response preferences

stanfordnlp/SHP

385K human preferences for answer helpfulness across 18 topics.

What's inside

SHP is a dataset of 385K collective human preferences between answers to questions and instructions. Topics span 18 domains, from cooking to legal advice; preferences reflect the helpfulness of one answer relative to another. The data is in English, in text and tabular modalities, designed for training models.

What you can do with it

  • RLHF/DPO fine-tuning on human preferences
  • Training reward models for answer helpfulness
  • Evaluating instruct models across 18 topics
Tasks
Text generation, Question answering
Languages
English
Size
100K–1M rows
Downloads this month
13.1K
Likes
333

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("stanfordnlp/SHP")