Bespoke Curator
A Python library for bulk inference and scalable synthetic-data curation, for preparing post-training datasets and extracting structured data from LLM outputs, with a viewer for monitoring generation.
Bespoke Curator makes it easier to build synthetic-data pipelines, whether for training a model or extracting structured data. It provides a Python library for generating and curating synthetic data with first-class support for structured outputs, a viewer for monitoring data as it's generated, and built-in performance optimizations for asynchronous operations.
It supports batch processing across OpenAI, Anthropic, Gemini and other compatible APIs, and can execute code that it generates through local multiprocessing, Ray, Docker or e2b backends. Fine-tuning integrations let curated data flow directly into a LoRA fine-tune, through the Tinker SDK or a managed Fireworks AI trainer, with the same trainer interface used to sample from the resulting model.
It has been used to build several published reasoning datasets, including OpenThoughts and Bespoke-Stratos, that have trended on Hugging Face.