FinetuneGemma3n

A small cookbook for fine-tuning Gemma-3N models using the Unsloth FastModel wrapper and TRL (SFTTrainer).

This repository contains a runnable Colab/VS Code notebook that demonstrates how to:

• load a Gemma-3N base model (4-bit quantized),
• prepare conversational datasets for instruction/response fine-tuning,
• apply LoRA-style PEFT adapters,
• train using TRL's SFTTrainer configured to train only on assistant responses,
• save the resulting adapters and export to formats such as merged fp16 or GGUF for deployment.

Quick links

• Notebook: finetuneGemma3n.ipynb (Colab badge available inside the notebook)

Requirements

• Python 3.8+ (recommended 3.10+)
• GPU with CUDA for training (optional for small tests)
• The notebook installs the required libraries when run in Colab. Locally, install the packages below:

pip install -r requirements.txt

(If there is no requirements.txt, the notebook installs the necessary packages automatically.)

Getting started (Colab)

1. Open finetuneGemma3n.ipynb in Colab using the badge at the top of the notebook.
2. Run the installation cells to install dependencies.
3. Edit the model/dataset cells as needed (e.g., change model name or dataset split).
4. Run the training cells.

Key notebook sections

• Installation: installs Unsloth and supporting libraries (bitsandbytes, accelerate, trl, peft, etc.)
• Load the Model: demonstrates FastModel.from_pretrained and 4-bit loading
• Dataset Preparation: loads a dataset, standardizes chat format and masks instruction tokens
• Training: configures TRL's SFTTrainer and trains only on assistant responses
• Saving: shows how to save LoRA adapters, merged fp16 model, and export GGUF

Saving and Export

• model.save_pretrained("gemma-3n") and tokenizer.save_pretrained("gemma-3n") saves LoRA adapters locally.
• The notebook contains examples to merge and save to fp16 for VLLM or export to GGUF for llama.cpp.

Troubleshooting

• Notebook rendering on GitHub: If you run into widget metadata rendering errors (missing metadata.widgets.state), open the notebook locally and remove metadata.widgets or run a small nbformat script to clean metadata. The repository's notebook has been cleaned for GitHub rendering.
• Out-of-memory: reduce batch size, use gradient accumulation, or use 4-bit loading as shown.