Developed by two Google Summer of Code students at DeepMind, Facet AI is a no-code web platform to manage the fine-tuning workflows of DeepMind's Gemma 3 family of multimodal models. Handling everything from dataset preprocessing to model export, Facet AI lets you focus on what matters for your use case, not writing boilerplate code and scrolling through pages of documentations and cookbook.
We cannot wait to see what you bring to the Gemmaverse with Facet AI!
Cookbooks and Colab notebooks are amazing, except that they don't scale when you have multiple datasets, models, and experiments to track. Facet AI not only provides an open-sourced backend to streamline and centralize all your fine-tuning needs, but also makes it easy for non-technical users to leverage the power of cutting-edge post-training techniques and Google Cloud without writing a single line of code. We're 100% IaC so it's extremely easy to deploy on your own cloud.
Watch our demo video!! 👇👇
| Category | Feature | Description | Supported Formats/Methods |
|---|---|---|---|
| Dataset | Preprocessing | Process vision and text data from custom uploads and Hugging Face | Custom files, Hugging Face datasets |
| Augmentation | Enhance datasets using NLP techniques and LLM-based synthetic generation | Transformers, Gemini, NLTK generation | |
| Training | Multiple Frameworks | Fine-tune using industry-standard frameworks | Hugging Face Transformers, Unsloth |
| Training Methods | Support for various post-training techniques | SFT (domain adaptation), DPO/ORPO (preference alignment), GRPO (reasoning tasks) | |
| Optimized Training | Flexible training approaches with quantization support | PEFT (LoRA, QLoRA), Full tuning, 4-bit/8-bit quantization | |
| Multimodal Support | Fine-tune both text and vision models | Text and multimodal datasets | |
| Utilities | Model Export | Export trained models to multiple deployment formats | Adapters, merged, quantized, GGUF, Hub or GCS |
| Training Monitoring | Comprehensive logging and monitoring to track your train jobs | Weights & Biases, trackio (coming soon) |
|
| Model Evaluation | Task and metric evaluation, batch inference testing | Benchmarks, custom dataset, "vibe check", model comparison | |
| Model Deployment | One-click deployment of trained models to GCP for inference | Cloud Run |
- More data augmentation for vision and text datasets
- Audio modality support for Gemma 3n
- Full feature plan is tracked in Project Roadmap
-
Use it for free on our website platform: https://gemma-facet.vercel.app/
-
Deploy the project to your own google cloud project with our Terraform and Cloud Build, see
infrastructure/README.mdfor details. -
Use this as a reference to build your own fine-tuning service since we're fully open source!
- User docs are live on here
- Developer docs are work in progress, use
README.mdin each directory for now - API reference is auto generated by FastAPI and available together with user docs.