DataDesigner

Why we included this project

When your training runs stall because you can't get enough varied, well-structured examples, DataDesigner is worth trying. It is a configurable pipeline rather than a thin wrapper around an LLM prompt: statistical samplers, seed data, and language models combine to produce datasets with the distributions and field correlations your workload actually needs. Validators written in Python or SQL (or your own custom checks) plus an LLM-as-a-judge scoring pass catch bad rows before they reach a downstream run, and preview mode lets you refine the recipe before committing tokens to a full generation. The async engine runs independent columns in parallel and adapts concurrency per provider and model, so most pipelines get faster without config changes. It sits between ad-hoc scripts and heavyweight data platforms, which is exactly where many teams end up.