Welcome to Gemma Fine-Tuning UI

This application allows you to fine-tune Google's Gemma models on your own datasets with a user-friendly interface.

Features:

• Upload and preprocess your datasets in various formats (CSV, JSONL, TXT)
• Configure model hyperparameters for optimal performance
• Visualize training progress in real-time
• Export your fine-tuned model in different formats

Getting Started:

1. Navigate to the Dataset Upload tab to prepare your data
2. Configure your model and hyperparameters in the Model Configuration tab
3. Start and monitor training in the Training tab
4. Export your fine-tuned model in the Export Model tab

For more details, check the Documentation tab.

Upload and prepare your dataset for fine-tuning

Select a model and configure hyperparameters

Train your model and monitor progress

Export your fine-tuned model

Gemma Fine-Tuning Documentation

Supported Models

This application supports fine-tuning the following Gemma models:

• google/gemma-2-2b-it
• google/gemma-2-9b-it
• google/gemma-2-27b-it

Dataset Format

Your dataset should follow one of these formats:

CSV

prompt,completion
"What is the capital of France?","The capital of France is Paris."
"How does photosynthesis work?","Photosynthesis is the process..."

JSONL

{"prompt": "What is the capital of France?", "completion": "The capital of France is Paris."}
{"prompt": "How does photosynthesis work?", "completion": "Photosynthesis is the process..."}

Plain Text

What is the capital of France?
###
The capital of France is Paris.
###
How does photosynthesis work?
###
Photosynthesis is the process...

Fine-Tuning Parameters

Basic Parameters

• Learning Rate: Controls how quickly the model adapts to the training data. Typical values range from 1e-5 to 5e-5.
• Batch Size: Number of samples processed before model weights are updated. Higher values require more memory.
• Number of Epochs: Number of complete passes through the training dataset. More epochs can lead to better results but may cause overfitting.
• Max Sequence Length: Maximum length of input sequences. Longer sequences require more memory.

Advanced Parameters

• Gradient Accumulation Steps: Accumulate gradients over multiple batches to simulate larger batch size.
• Warmup Steps: Number of steps for learning rate warmup. Helps stabilize training in the early phases.
• Weight Decay: L2 regularization factor to prevent overfitting.
• LoRA Parameters: Controls the behavior of LoRA (Low-Rank Adaptation), a parameter-efficient fine-tuning technique.

Export Formats

• PyTorch: Standard PyTorch model format (.pt or .bin files with model architecture).
• GGUF: Compact format optimized for efficient inference (especially with llama.cpp).
• Safetensors: Safe format for storing tensors, preventing arbitrary code execution.

Quantization

Quantization reduces model size and increases inference speed at the cost of some accuracy:

• None: No quantization, full precision (usually FP16 or BF16).
• int8: 8-bit integer quantization, good balance of speed and accuracy.
• int4: 4-bit integer quantization, fastest but may reduce accuracy more significantly.