README.md

---
title: README
emoji: 🍫
colorFrom: yellow
colorTo: green
sdk: static
pinned: false
---

<p align="center">
    <img src="https://raw.githubusercontent.com/ml6team/fondant/main/docs/art/fondant_banner.svg" height="250px"/>
</p>
<p align="center">
    <i>Sweet data-centric foundation model fine-tuning</i>
    <br>
    <a href="https://fondant.readthedocs.io/en/stable/"><strong>Explore the docs »</strong></a>
    <br>
    <br>
    <a href="https://discord.gg/HnTdWhydGp"><img alt="Discord" src="https://dcbadge.vercel.app/api/server/HnTdWhydGp?style=flat-square"></a>
</p>

---
**Fondant helps you create high quality datasets to train or fine-tune foundation models such as:**

- 🎨 Stable Diffusion  
- 📄 GPT-like Large Language Models (LLMs)  
- 🔎 CLIP  
- ✂️ Segment Anything (SAM)  
- ➕ And many more

## 🪤 Why Fondant?

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.

<p align="right">(<a href="#chocolate_bar-fondant">back to top</a>)</p>