Dataset Card for RLHF

This multi-turn dialogue dataset is annotated based on different evaluation criteria of helpfulness and harmlessness:

Preference

The helpfulness is from instructGPT and the harmlessness is from BeaverTails.

You can find more information in these papers:

instructGPT:https://arxiv.org/pdf/2203.02155

BeaverTails:https://arxiv.org/abs/2307.04657

Helpfulness	Harmlessness
Writing in clear language	Animal Abuse
Answering the question they meant to ask, even if they mis-asked it	Child Abuse
Being sensitive to internationality	Controversial Topics, Politics
If an instruction is too confusing, asking for clarification and explaining why the instruction is confusing	Discrimination, Stereotype, Injustice
Not giving overly long or rambling answers, or repeating information from the question	Drug Abuse, Weapons, Banned Substance
Not assuming extraneous extra context outside of what’s given (besides things like facts about the world), unless that’s an implied part of the task	Financial Crime, Property Crime, Theft
	Hate Speech, Offensive Language
	Misinformation Regarding ethics, laws, and safety
	Non-Violent Unethical Behavior
	Privacy Violation
	Self-Harm
	Sexually Explicit, Adult Content
	Terrorism, Organized Crime
	Violence, Aiding and Abetting, Incitement

Usage

The code snippet below demonstrates how to load this dataset:

from datasets import load_dataset

# Load the whole dataset
dataset = load_dataset('NJU-Alignment/RLHF')

Papers

In production......

Contact

please contact NJU-Alignment Team at Nanjing University for the whole dataset.

Warning: this dataset contains data that may be offensive or harmful. The data are intended for research purposes. The views expressed in the data do not reflect the views of NJU-Alignment Team or any of its members.