Capability · Framework — fine-tuning
Distilabel
Distilabel lets you build data pipelines where each step is an LLM or a transform. You can distil from a strong teacher model, generate preference pairs for DPO/ORPO, run UltraFeedback-style scoring, or synthesise instruction datasets. It integrates with Argilla for human review and Hugging Face datasets for storage.
Framework facts
- Category
- fine-tuning
- Language
- Python
- License
- Apache-2.0
- Repository
- https://github.com/argilla-io/distilabel
Install
pip install 'distilabel[vllm,openai,anthropic]'Quickstart
from distilabel.pipeline import Pipeline
from distilabel.steps import LoadDataFromHub
from distilabel.steps.tasks import TextGeneration
from distilabel.llms import OpenAILLM
with Pipeline('demo') as pipeline:
load = LoadDataFromHub(repo_id='argilla/10k-questions')
generate = TextGeneration(llm=OpenAILLM(model='gpt-4o-mini'))
load >> generate
distiset = pipeline.run(parameters={load.name: {'split': 'train'}})
distiset.push_to_hub('me/my-synthetic-ft')Alternatives
- Augmentoolkit — long-context synthetic data
- Bonito — instruction-tuning data generator
- PromptWright — simpler synthetic gen
- OpenAI Evals dataset tooling
Frequently asked questions
What's Distilabel best at?
Preference-pair generation for DPO/ORPO and distillation at scale. The DAG model makes it straightforward to combine generation + critique + judge steps.
Does it support local models?
Yes — vLLM, Ollama, LiteLLM, and Hugging Face TGI are supported alongside OpenAI, Anthropic, and Mistral APIs.
Sources
- Distilabel — GitHub — accessed 2026-04-20
- Distilabel — docs — accessed 2026-04-20