Fine-tuning Gemma Models
Transform Gemma models into specialized AI assistants tailored to your specific needs. Facet makes fine-tuning accessible with a no-code interface, powerful training algorithms, and optimized configurations.What You Can Do
Multiple Model Sizes
Fine-tune Gemma 3 models from 270M to 27B parameters
Various Training Methods
Use SFT, DPO, ORPO, and GRPO for different objectives
Efficient Training
LoRA and QLoRA for memory-efficient fine-tuning
Vision Support
Fine-tune multimodal models with text and images
Supported Models
Facet supports the complete Gemma 3 family:- Gemma 3 270M: Fast, lightweight for simple tasks
- Gemma 3 1B: Balanced performance and efficiency
- Gemma 3 4B: High-quality responses, good for most tasks
- Gemma 3 12B: Advanced reasoning and complex tasks
- Gemma 3 27B: State-of-the-art performance
Training Methods
Choose the right training method for your objective:Supervised Fine-Tuning (SFT)
Best for: General conversation, instruction following, domain adaptation- Uses your processed dataset with conversations
- Learns from human demonstrations
- Most common fine-tuning approach
Direct Preference Optimization (DPO)
Best for: Aligning models with human preferences- Uses paired examples (chosen vs rejected responses)
- Learns to prefer better responses
- Great for safety and helpfulness alignment
Odds Ratio Preference Optimization (ORPO)
Best for: Alternative preference learning approach- Similar to DPO but with different optimization
- Can be more stable in some cases
- Good alternative to DPO
Group-Related Policy Optimization (GRPO)
Best for: Reasoning tasks, math problems, structured thinking- Uses reward functions to score responses
- Learns to maximize rewards through reasoning
- Perfect for complex problem-solving tasks
Training Efficiency Options
Full Fine-tuning
- Updates all model parameters
- Requires more memory and compute
- Best for large datasets and major domain changes
LoRA (Low-Rank Adaptation)
- Updates only a small subset of parameters
- Much more memory efficient
- Good for most fine-tuning tasks
QLoRA (Quantized LoRA)
- Combines quantization with LoRA
- Most memory efficient option
- Recommended for most use cases
Start with QLoRA for most tasks. It provides excellent results while using
minimal resources.
Getting Started
1
Prepare Your Dataset
Upload and process your data using the dataset preprocessing guide.
Make sure your dataset is processed and ready before starting training.
2
Choose Your Model
Select the Gemma model size that fits your needs and resources.
Larger models generally perform better but require more compute resources.
3
Configure Training
Set up your training parameters, method, and efficiency options.
Use the recommended settings for your first training run, then experiment
with custom parameters.
4
Start Training
Launch your training job and monitor progress.
Training can take several hours depending on your dataset size and model
choice.
5
Evaluate Results
Test your fine-tuned model and review performance metrics.
Use the evaluation tools to assess your model’s performance on test data.
Training Configuration
Essential Settings
- Base Model: Choose from available Gemma models
- Training Method: SFT, DPO, ORPO, or GRPO
- Efficiency Method: Full, LoRA, or QLoRA
- Dataset: Your processed dataset
- Hyperparameters: Learning rate, batch size, epochs
Advanced Options
- Evaluation: Set up validation and testing
- Monitoring: Track training with Weights & Biases
- Export: Configure model export formats
- Reward Functions: Set up graders for GRPO training
Best Practices
Next Steps
Once your model is trained:- Evaluate Performance: Use the evaluation tools to assess your model
- Export Your Model: Download in various formats (adapter, merged, GGUF)
- Deploy: Set up your model for production use
- Iterate: Use feedback to improve your model with additional training