Distilabel

Distilabel, created by the Argilla team, is a Python framework for generating synthetic datasets and AI feedback through composable pipelines that implement methods from published research, such as preference ranking and instruction evolution. A unified API abstracts over many LLM backends, including OpenAI, Anthropic, Cohere, Groq, Mistral, Vertex AI, and local options like Ollama, vLLM, llama-cpp, and Transformers, so the same pipeline can switch models freely. Pipelines are scalable and fault tolerant, with Ray available for distributed execution and Outlines or Instructor for structured generation. The framework has been used to build public datasets such as the OpenHermes preference set and the distilabeled Intel Orca DPO pairs. It requires Python 3.9 or later and is open source under the Apache 2.0 license, now community maintained. Engineers who need labeled training data, preference pairs for DPO or RLHF, or automated quality judgments over existing corpora are its main users.