Curator

Bespoke Labs built Curator for the synthetic data loops behind post-training: Python code defines prompt and parse steps, and the library fans them out across LLM providers with asynchronous batching, automatic retries, caching, and fault recovery, so a large generation job survives interruptions and does not repeat paid calls. Structured output is first class through Pydantic models, and backends include OpenAI, Anthropic, Gemini, and DeepSeek directly, batch APIs for the major providers, LiteLLM for dozens more, and local inference through vLLM or Ollama for fully offline runs, alongside a managed fine-tuning path on Fireworks AI. A companion Curator Viewer displays generated data streaming in for quick quality checks. The library produced the OpenThoughts-114k and OpenThoughts2-1M reasoning datasets used to train the OpenThinker models, a solid demonstration of its scale. Installation is pip install bespokelabs-curator, documentation sits at docs.bespokelabs.ai, and the code is Apache 2.0 with about 1,700 GitHub stars and active development through 2026. Teams reach for it when ad hoc scripts for dataset generation stop scaling.