Sweet data-centric foundation model fine-tuning
Explore the docs »
Fondant helps you create high quality datasets to fine-tune foundation models such as:
- 🎨 Stable Diffusion
- 📄 GPT-like Large Language Models (LLMs)
- 🔎 CLIP
- ✂️ Segment Anything (SAM)
- ➕ And many more
Foundation models simplify inference by solving multiple tasks across modalities with a simple prompt-based interface. But what they've gained in the front, they've lost in the back. These models require enormous amounts of data, moving complexity towards data preparation, and leaving few parties able to train their own models.
We believe that innovation is a group effort, requiring collaboration. While the community has been building and sharing models, everyone is still building their data preparation from scratch. Fondant is the platform where we meet to build and share data preparation workflows.
Fondant offers a framework to build composable data preparation pipelines, with reusable components, optimized to handle massive datasets. Stop building from scratch, and start reusing components to:
- Extend your data with public datasets
- Generate new modalities using captioning, segmentation, translation, image generation, ...
- Distill knowledge from existing foundation models
- Filter out low quality data
- Deduplicate data
And create high quality datasets to fine-tune your own foundation models.
Curious to see what Fondant can do? Have a look at our example pipelines:
Our example pipeline to generate data for ControlNet fine-tuning allows you to create models that you can control using inpainting, segmentation, and regeneration. All you need to get started is a set of prompts describing the type of images to generate.
For instance, using our ControlNet model fine-tuned on interior design images, allows you to generate the room of your dreams:
| Input image | Output image |
|---|---|
 |
 |
Want to try out the resulting model yourself, head over to our Hugging Face space!
Using our example pipeline to fine-tune Stable Diffusion allows you to create models that generate better images within a specific domain. All you need to get started is a small seed dataset of example images.
Eg. generating logos:
| Stable Diffusion 1.5 | Fine-tuned Stable Diffusion 1.5 |
|---|---|
 |
 |
Fondant comes with a library of reusable components, which can jumpstart your pipeline.
| Component | Description |
|---|---|
| load_from_hf_hub | Load a dataset from the Hugging Face Hub |
| image_embedding | Create embeddings for images using a model from the HF Hub |
| embedding_based_laion_retrieval | Retrieve images-text pairs from LAION using embedding similarity |
| prompt_based_laion_retrieval | Retrieve images-text pairs from LAION using prompt similarity |
| download_images | Download images from urls |
| image_resolution_filtering | Filter images based on their resolution |
| caption images | Generate captions for images using a model from the HF Hub |
| segment_images | Generate segmentation maps for images using a model from the HF Hub |
Fondant can be installed using pip:
pip install fondant
For the latest development version, you might want to install from source instead:
pip install git+https://github.com/ml6team/fondant.git
Fondant is flexible and works by leveraging kubeflow pipelines on any kubernetes cluster. All fondant needs is an url pointing to the kubeflow pipeline host and a Object Storage provider (S3, GCS, etc) to store data produced in the pipeline between steps.
We have compiled some references and created some scripts to get you started with setting up the required infrastructure.
Fondant allows you to easily define data pipelines comprised of both reusable and custom
components. The following pipeline for instance uses the reusable load_from_hf_hub component
to load a dataset from the Hugging Face Hub and process it using a custom component:
from fondant.pipeline import ComponentOp, Pipeline, Client
def build_pipeline():
pipeline = Pipeline(pipeline_name="example pipeline", base_path="fs://bucket")
load_from_hub_op = ComponentOp.from_registry(
name="load_from_hf_hub",
arguments={"dataset_name": "lambdalabs/pokemon-blip-captions"},
)
pipeline.add_op(load_from_hub_op)
custom_op = ComponentOp(
component_spec_path="components/custom_component/fondant_component.yaml",
arguments={
"min_width": 600,
"min_height": 600,
},
)
pipeline.add_op(custom_op, dependencies=load_from_hub_op)
return pipeline
if __name__ == "__main__":
client = Client(host="https://kfp-host.com/")
pipeline = build_pipeline()
client.compile_and_run(pipeline=pipeline)To create a custom component, you first need to describe its contract as a yaml specification. It defines the data consumed and produced by the component and any arguments it takes.
name: Custom component
description: This is a custom component
image: custom_component:latest
consumes:
images:
fields:
data:
type: binary
produces:
captions:
fields:
data:
type: utf8
args:
argument1:
description: An argument passed to the component at runtime
type: str
argument2:
description: Another argument passed to the component at runtime
type: strOnce you have your component specification, all you need to do is implement a single .transform
method and Fondant will do the rest. You will get the data defined in your specification as a
Dask dataframe, which is evaluated lazily.
from fondant.component import TransformComponent
class ExampleComponent(TransformComponent):
def transform(self, dataframe, *, argument1, argument2):
"""Implement your custom logic in this single method
Args:
dataframe: A Dask dataframe containing the data
argumentX: An argument passed to the component
"""Fondant is currently in the alpha stage, offering a minimal viable interface. While you should expect to run into rough edges, the foundations are ready and Fondant should already be able to speed up your data preparation work.
The following topics are on our roadmap
- Local pipeline execution
- Non-linear pipeline DAGs
- LLM-focused example pipelines and reusable components
- Static validation, caching, and partial execution of pipelines
- Data lineage and experiment tracking
- Distributed execution, both on and off cluster
- Support other dataframe libraries such as HF Datasets, Polars, Spark
- Move reusable components into a decentralized component registry
- Create datasets of copy-right free data for fine-tuning
- Create reusable components for bias detection and mitigation
The roadmap and priority are defined based on community feedback. To provide input, you can join our discord or submit an idea in our Github Discussions.
For a detailed view on the roadmap and day to day development, you can check our github project board.
We welcome contributions of different kinds:
| Issues | If you encounter any issue or bug, please submit them as a Github issue. You can also submit a pull request directly to fix any clear bugs. |
| Suggestions and feedback | If you have any suggestions or feedback, please reach out via our Discord server or Github Discussions! |
| Framework code contributions | If you want to help with the development of the Fondant framework, have a look at the issues marked with the good first issue label. If you want to add additional functionality, please submit an issue for it first. |
| Reusable components | Extending our library of reusable components is a great way to contribute. If you built a component which would be useful for other users, please submit a PR adding them to the components/ directory. |
| Example pipelines | If you built a pipeline with Fondant which can serve as an example to other users, please submit a PR adding them to the examples/ directory. |
We use poetry and pre-commit to enable a smooth developer flow. Run the following commands to set up your development environment:
pip install poetry
poetry install
pre-commit install