distilabel

// Overview

distilabel is a Python framework for generating synthetic data and AI feedback. You write pipelines in code that call large language models to create new examples, label existing ones, and judge their quality, then export the result as a dataset you can use for fine-tuning or evaluation.

It is aimed at engineers and researchers who need to build training and preference datasets without labeling everything by hand. It works for traditional NLP tasks like classification and extraction, as well as generative tasks like instruction following, dialogue, and LLM-as-a-judge scoring.

As a synthetic-data and labeling tool, distilabel sits between your raw inputs and your model training. It uses one unified API across many LLM providers (OpenAI, Anthropic, Mistral, Hugging Face, Ollama, vLLM, and others), so you can swap models without rewriting your pipeline.

// What it does

• Pipeline API that chains steps and tasks such as TextGeneration to produce and process data
• One unified interface to many LLM backends, including OpenAI, Anthropic, Mistral, Cohere, Groq, Ollama, vLLM, and Hugging Face Inference Endpoints
• Built-in tasks for synthetic data generation, labeling, and AI feedback (LLM-as-a-judge)
• Optional extras let you install only the LLM integrations you need (for example, hf-inference-endpoints, openai, vllm)
• Integrates with the Hugging Face datasets ecosystem for loading inputs and pushing results to the Hub
• Designed for scalable pipelines with fault tolerance, based on published research methods

// Getting started

Install distilabel with the LLM integration you want, then define a Pipeline and run it over a dataset. The example below uses Hugging Face Inference Endpoints.

pip install "distilabel[hf-inference-endpoints]" --upgrade

from datasets import load_dataset

from distilabel.models import InferenceEndpointsLLM
from distilabel.pipeline import Pipeline
from distilabel.steps.tasks import TextGeneration

with Pipeline() as pipeline:
    TextGeneration(
        llm=InferenceEndpointsLLM(
            model_id="meta-llama/Meta-Llama-3.1-8B-Instruct",
            generation_kwargs={"temperature": 0.7, "max_new_tokens": 512},
        ),
    )

if __name__ == "__main__":
    dataset = load_dataset("distilabel-internal-testing/instructions", split="test")
    distiset = pipeline.run(dataset=dataset)
    distiset.push_to_hub(repo_id="distilabel-example")

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Generate synthetic instruction or dialogue datasets to fine-tune your own LLM
• Build preference (DPO) datasets by having an LLM judge and rank candidate responses
• Label or filter existing data at scale using AI feedback instead of manual annotation
• Create task-specific evaluation sets that follow methods from recent research papers

// How distilabel compares

distilabel alongside other open-source synthetic data tools AI/TLDR tracks, ranked by GitHub stars.