About Reinforcement Loving

What This Is

Reinforcement Loving is a public dataset of human judgments about text and AI responses.

People rate responses, compare different answers, answer short judgment questions, write responses of their own, and review longer documents.

Those contributions are stored as training and evaluation data that can be downloaded and used with language models.

What People Do

The review feed contains several kinds of tasks.

You may be asked to rate a response, compare two to four responses to the same prompt, say how strongly you agree with a statement, write an answer yourself, or review a longer document.

Some questions have fairly clear answers. Others involve tradeoffs where people may reasonably disagree.

The dataset records those judgments, including disagreement and uncertainty.

What Is In The Dataset

The dataset includes:

• response ratings
• comparisons between two to four responses
• agreement ratings on short statements and questions
• written responses
• document reviews

These are stored as different kinds of data rather than being combined into one score.

For example, a five point agreement rating is kept separate from a preference between two answers.

Reviewer identities are anonymized in downloadable data. Contributions from the same reviewer can be associated with one another without identifying that person.

Why The Data Is Collected

Language models are often trained or evaluated using examples of preferred and nonpreferred behavior.

Reinforcement Loving collects human judgments that can be used for that purpose.

These judgments may concern factual quality, usefulness, honesty, safety, fairness, consideration for other people, or other aspects of a response.

The dataset does not assume that every question has one correct judgment. Different reviewers may disagree.

What Can Be Submitted

People can submit prompts, responses, documents, and other material that can be reviewed through the site.

Material does not need to be good in order to be useful. Weak responses, neutral material, close comparisons, and disputed cases can all produce useful ratings.

Submitted material must comply with the Acceptable Use Policy.

Illegal content, private personal information, and other prohibited material is not permitted.

Using The Data

The dataset can be downloaded for research, fine tuning, evaluation, classification, or other work with language models.

The downloadable data includes the review type, the material that was reviewed where redistribution is permitted, the judgment that was recorded, and related metadata.