EasyDistill 2
A config-driven knowledge-distillation toolkit that turns a black-box teacher model into training data, composing generation, evaluation, filtering, rewriting, balancing and preference-scoring operators into pipelines that build SFT and DPO datasets.
EasyDistill 2 builds supervised fine-tuning datasets and chosen/rejected preference pairs for alignment, from text instructions, chain-of-thought reasoning, vision-language conversations, agent trajectories and text-to-image data. Generation works against any OpenAI-compatible endpoint, including OpenAI's own API, Azure OpenAI, vLLM, local servers, and Alibaba Cloud's PAI-Token and PAI-EAS.
A single CLI entry point runs every job; a job_type field in its config selects the workflow, and each workflow ships paired configs for the PAI-Token and PAI-EAS backends. easydistill --list-jobs and easydistill --list-models list the supported jobs and the model zoo without needing a config file first.
Exported formats work directly with LLaMA-Factory and ms-swift. It is a full rewrite of an earlier 1.0 release and is not backward compatible with 1.0 configs or APIs, though the 1.0 implementation remains available on its own branch and tag.