CA‑1 Human Preference Alignment Dataset

*This dataset supports online use.Click here to jump.

CA‑1, short for Collective Alignment 1, was released by OpenAI in 2025. It focuses on human value judgments and preferences on the default behaviors of AI models. It is a human feedback behavior dataset that combines model-generated content and annotator evaluations.

The dataset consists of 1,078 prompt comparison items and 18,384 evaluations provided by 1,012 annotators, along with demographic information and scoring rationale. It is suitable for studying group alignment differences, guiding model behavior norms, and developing value-sensitive reward mechanisms.

Data composition:

• Prompt Comparisons: Each item contains a synthetic prompt and four corresponding candidate responses (A–D), which are used to compare the differences in value preferences among different responses.
• Annotators: This records the basic demographic information of each annotator, including age, gender, education level, and country of origin, for analyzing the differences in value preferences among different groups of people.
• Merged Comparisons–Annotators: This combines prompts, candidate responses, and annotators’ evaluations, along with contextual information about the annotators, to create a complete prompt–response–evaluation–demographics map.