curator

Why we included this project

Most teams find assembling post-training data is where fine-tuning actually gets hard, and Curator targets that step directly. You write a plain Python function that calls a model, and the library runs those calls across thousands of rows for you, handling batching, caching, and fault recovery in the background, so generating reasoning traces, instruction pairs, or structured JSON doesn't mean babysitting a long job. The same pipeline doubles as structured data extraction, so one interface serves both a team preparing training data and a team pulling JSON out of messy documents. If you're distilling a strong model into a smaller one or putting together an SFT or RLHF set, this is a solid practical starting point.