ml6team/fondant

Large-scale data processing made easy and reusable
Explore the docs »

🍫Fondant is an open-source framework that simplifies and speeds up data processing development through reusable components

It offers:

• 🔧 Plug ‘n’ play composable pipelines for creating datasets for
  • AI image generation model fine-tuning (Stable Diffusion, ControlNet)
  • Large language model fine-tuning (LLaMA, Falcon)
  • Code generation model fine-tuning (StarCoder)
• 🧱 Library of off-the-shelf reusable components for
  • Extracting data from public sources such as Common Crawl, LAION, ...
  • Filtering on
    • Content, e.g. language, visual style, topic, format, aesthetics, etc.
    • Context, e.g. copyright license, origin
    • Metadata
  • Removal of unwanted data such as toxic, NSFW or generated content
  • Removal of unwanted data patterns such as societal bias
  • Transforming data (resizing, cropping, reformatting, …)
  • Tuning the data for model performance (normalization, deduplication, …)
  • Enriching data (captioning, metadata generation, synthetics, …)
  • Transparency, auditability, compliance
• 📖 🖼️ 🎞️ ♾️ Out of the box multimodal capabilities: text, images, video, etc.
• 🐍 Standardized, Python/Pandas-based way of creating custom components
• 🏭 Production-ready, scalable deployment
• ☁️ Multi-cloud integrations

🪤 Why Fondant?

In the age of Foundation Models, control over your data is key and building pipelines for large-scale data processing is costly, especially when they require advanced machine learning-based operations. This need not be the case, however, if processing components would be reusable and exchangeable and pipelines were easily composable. Realizing this is the main vision behind Fondant.

(back to top)

💨 Getting Started

Anxious to get started? Here's is a step by step guide to get your first pipeline up and running.

🪄 Example pipelines

Curious to see what Fondant can do? Have a look at our example pipelines:

Filtering creative commons image dataset

We have published an image dataset containing 25 million images. As a result, we have provided a sample pipeline that demonstrates the download and filtering of these images. In the pipeline folder, you will find detailed instructions on how to execute the pipeline and explore the images.

Fine-tuning ControlNet

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!

Fine-tuning Stable Diffusion

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

Training Starcoder

Using our example pipeline to train StarCoder provides a starting point to create datasets for training code assistants.

(back to top)

🧩 Reusable components

Fondant comes with a library of reusable components, which can jumpstart your pipeline.

(back to top)

⚒️ Installation

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

🧱 Running Fondant pipelines

There are 3 ways to run fondant pipelines:

• Local runner: leverages docker compose. The local runner is mainly aimed at helping you develop fondant pipelines and components faster since it allows you to develop on your local machine or a Virtual Machine.
• This enables you to quickly iterate on development.Once you have a pipeline developed, you can use the other runners mentioned below for better scaling, monitoring and reproducibility.
• Vertex runner: Uses Google cloud's Vertex AI pipelines to help you orchestrate your Fondant pipelines in a serverless manner. This makes it easy to scale up your pipelines without worrying about infrastructure deployment.
• Kubeflow runner: Leverages Kubeflow pipelines on any Kubernetes cluster. All Fondant needs is a url pointing to the Kubeflow pipeline host and an 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.

It is worth noting that the same pipeline can be used across all runners allowing you to quickly develop and iterate using the local runner and then using the Vertex or Kubeflow runner to run a large scale pipeline.

(back to top)

👨‍💻 Usage

Pipeline

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:

pipeline.py

from fondant.pipeline import ComponentOp, 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_dir="components/custom_component",
    arguments={
        "min_width": 600,
        "min_height": 600,
    },
)
pipeline.add_op(custom_op, dependencies=load_from_hub_op)

Component

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: str

Once you have your component specification, all you need to do is implement a constructor and a single .transform method and Fondant will do the rest. You will get the data defined in your specification partition by partition as a Pandas dataframe.

import pandas as pd
from fondant.component import PandasTransformComponent


class ExampleComponent(PandasTransformComponent):

    def __init__(self, *args, argument1, argument2) -> None:
        """
        Args:
            argumentX: An argument passed to the component
        """
        # Initialize your component here based on the arguments

    def transform(self, dataframe: pd.DataFrame) -> pd.DataFrame:
        """Implement your custom logic in this single method
        Args:
            dataframe: A Pandas dataframe containing the data
        Returns:
            A pandas dataframe containing the transformed data
        """

For more advanced use cases, you can use the DaskTransformComponent instead.

Running your pipeline

Once you have a pipeline you can easily run (and compile) it by using the built-in CLI:

fondant run local pipeline.py

To see all available runner and arguments you can check the fondant CLI help pages

fondant --help

Or for a subcommand:

fondant <subcommand> --help

(back to top)

🚧 Current state and roadmap

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

• Non-linear pipeline DAGs
• LLM-focused example pipelines and reusable components
• 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.

(back to top)

👭 Contributing

We welcome contributions of different kinds

Environment setup

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

(back to top)