stanfordnlp/SHP

Dataset Overview

Dataset Name

Stanford Human Preferences Dataset (SHP)

Dataset Size

385K entries

Dataset Task Categories

• Text Generation
• Question Answering

Dataset Tags

• Human Feedback
• RLHF
• Preference
• Reddit
• Preference Model
• RL
• NLG
• Evaluation

Dataset Language

English

Dataset Content

SHP contains 385K human preference records for question/instruction answering, covering 18 domains such as cooking, legal advice, etc. Each example includes a Reddit post, a question/instruction, and two top comments, with one comment preferentially chosen by the Reddit community.

Dataset Structure

The dataset is divided into 18 directories, each representing a sub‑forum, containing JSONL files for training, validation, and testing.

Dataset Uses

For training RLHF reward models and NLG evaluation models.

Dataset Differences from Other Datasets

• Compared with the Anthropics HH‑RLHF dataset, SHP’s data are naturally occurring and human‑written, whereas HH‑RLHF’s responses are machine‑generated.
• Compared with the ELI5 dataset, SHP infers preferences using timestamps, while ELI5 only provides comments and scores.

Dataset Preprocessing

Preprocessing is minimal, limited to expanding sub‑forum‑specific abbreviations and removing URLs from hyperlinks.

Building Preference Models

It is recommended to fine‑tune large models, such as FLAN‑T5‑xl, to predict human preferences, and to report performance curves based on score_ratio.

Dataset Limitations

• SHP is not intended for harm reduction and does not contain toxic content needed for toxicity detection training.
• More preferred responses are not necessarily more factually accurate.

Dataset License

Data were scraped in accordance with Reddit API terms of use; user‑generated content remains owned by the users, and Reddit grants a non‑exclusive, non‑transferable, non‑sublicensable, revocable license.

Contact Information

Email: kawin@stanford.edu

Dataset Creators

Kawin Ethayarajh, Heidi (Chenyu) Zhang, Yizhong Wang, Dan Jurafsky