Curator

Curator is a Python library for developers building LLM training datasets or extracting structured records from existing data. It supports pipelines that generate data in stages and process many model requests.

• Structured data Define the fields model responses should contain and parse them into records for tasks such as sentiment analysis and product feature extraction.
• Synthetic datasets Chain generation stages to build training examples, including reasoning traces for math, coding and problem solving.
• Bulk inference Parallel requests, retries and response caching support large generation jobs, with provider batch APIs available for OpenAI, Anthropic, Gemini and Mistral.
• Model backends Run inference through Ollama or vLLM on your own hardware, or connect to OpenAI-compatible APIs and providers through LiteLLM.
• Fine-tuning Tinker and Fireworks AI integrations support LoRA training; Fireworks requires an account and runs training on its servers using uploaded chat data.

Curator is open source under Apache 2.0. Local backends let you use models in your own environment, while cloud API calls send prompts to the selected provider. It collects anonymized usage telemetry that you can disable.

The optional hosted viewer uploads generated datasets to Bespoke Labs for inspection. By default, anyone with the link can access those datasets; associating them with a Bespoke Labs account supports private access and collaborator sharing.