Get the 2026 ML Training Cookbook | 52 recipes — GRPO, Flow Matching, World Models, and everything in between Download Now →
Constitutional Generation
Generate synthetic training data that respects predefined constraints and quality rubrics — ensuring generated examples meet safety, format, and content standards without manual review. Constitutional generation produces synthetic data by following a written constitution or rubric that specifies constraints on the output. The rubric defines acceptable content, format requirements, safety boundaries, and quality criteria. The generating model receives the rubric as a system prompt when creating each example, and a judge model verifies compliance after generation. This produces training data that is safe, on-format, and high-quality without humans in the loop.
Papers, code, and datasets
Want a custom dataset for this?
We build high-quality synthetic and curated datasets for constitutional generation and other training techniques. Tell us about your data needs.