distilabel

Why we included this project

Distilabel is for teams that need training data for fine-tuning but do not want to hand-label thousands of examples. You assemble data-generation and AI-judging steps into Python pipelines and synthesize instruction pairs, preference data, and other training material at scale, following published research methods for tasks like preference labeling and reward-model data. A single API reaches across many model providers, so you are not tied to one vendor when generating or critiquing data. The project has shipped real results, including preference datasets with millions of examples and fine-tuned models that improved after AI-feedback filtering. If your bottleneck is clean, varied training data rather than model code, this framework is worth a close look.