Curator

Bespoke Curator is an open source Python library from Bespoke Labs for building synthetic data pipelines. It is used to generate and curate data for post training language models and to extract structured information from unstructured inputs, with the library handling the orchestration of large numbers of model calls. It is written for machine learning engineers and researchers who need training or evaluation data at scale. Key capabilities include first class support for structured outputs, asynchronous execution, caching and automatic recovery so interrupted runs resume without repeating completed work, and a wide range of inference options through LiteLLM, vLLM and popular batch APIs, which can reduce the cost of large jobs. An interactive Curator Viewer lets users monitor and explore a dataset while it is being generated. The documentation includes data curation recipes, fine tuning examples and an API reference. Bespoke Labs lists verifiers, Monte Carlo Tree Search over reasoning trajectories, data versioning and data quality indicators as areas of active development. Curator is free to use under the Apache 2.0 license. Users pay only for the model inference they run through their chosen providers, and Bespoke Labs does not publish a paid tier for the library. The same team publishes the Bespoke MiniCheck fact checking models, Bespoke MiniChart and the OpenThinker reasoning models, all documented alongside Curator on the same site.