About Safe Training Data

What This Is

Safe Training Data is an open dataset of text documents rated by people. The goal is to produce high-quality training data for building AI systems that are safe, honest, and beneficial.

Most publicly available training data is selected for size, coverage, or convenience. This dataset is different. It is curated specifically to support AI alignment—training models that behave in ways that are helpful to people and avoid causing harm.

How It Works

Anyone can upload a text document. The community then rates each document on a scale from unsafe to great. Over time, these ratings produce a dataset with clear, human-generated labels on what is beneficial, what is neutral, and what is harmful.

The result is a dataset you can use directly to fine-tune a language model, or as a reference set for classifying much larger collections of text. If you have a million documents and need to identify which ones are prosocial, a well-rated corpus gives you something to measure against.

Why This Matters

The behavior of a language model depends heavily on what it was trained on. Data that is thoughtful, constructive, and honest tends to produce models that are thoughtful, constructive, and honest. The reverse is also true.

By building a dataset where humans have explicitly identified what is beneficial, we provide a resource for anyone working on AI safety—whether they are fine-tuning a model directly, doing alignment research, or filtering a larger corpus for quality.

What Gets Uploaded

We welcome all kinds of content. The dataset benefits from having the full range—documents rated as beneficial, neutral, and harmful all serve a purpose. The rating system is what creates the value, not restrictions on what can be submitted.

That said, content must still comply with our Acceptable Use Policy. Illegal content, personal information, and similar material is not permitted.

Open and Free

The dataset is open. You can download it, use it for research, use it to train models, or build on it however you see fit. The ratings, the documents, and the metadata are all available.