README.md

metadata

title: FoodExtract v1
emoji: 🍕
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 5.0.0
app_file: app.py
pinned: false
license: mit
python_version: '3.11'
short_description: Extract food & drinks from text with Gemma 3

🍕 Fully Fine-Tune a Small Language Model (Gemma 3 270M)

A tutorial on how to fully fine-tune Google's Gemma 3 270M model using Hugging Face libraries to extract food and drink items from text.

📖 Overview

This project demonstrates Supervised Fine-Tuning (SFT) of a Small Language Model (SLM) for a specific task: extracting food and drink items from text. The fine-tuned model can process text inputs and return structured data about food/drink content.

Why Fine-tune a Small Language Model?

• ✅ Own the model - Run anywhere without API costs
• ✅ Simple tasks work well with smaller models
• ✅ No API calls needed - Run offline
• ✅ Batch processing - Much faster than API calls
• ✅ Task-specific optimization - Better performance on your use case

🎯 What We're Building

A model that extracts food and drink items from text, returning structured output:

Input:

A plate of rice cakes, salmon, cottage cheese and small cherry tomatoes with a cup of tea.

Output:

food_or_drink: 1
tags: fi
foods: rice cakes, salmon, cottage cheese, cherry tomatoes
drinks: cup of tea

🛠️ Technologies Used

• Model: Gemma 3 270M
• Dataset: FoodExtract-1k
• Libraries:
  • transformers - Model loading and inference
  • trl - Transformers Reinforcement Learning (SFT)
  • datasets - Data loading
  • accelerate - Training acceleration
  • gradio - Interactive demo

📋 Requirements

• Python 3.8+
• GPU with at least 16GB VRAM (Google Colab T4 works!)
• Hugging Face account (for uploading model)

🚀 Quick Start

1. Open in Google Colab

2. Enable GPU Runtime

• Go to Runtime → Change runtime type → Select GPU

3. Run All Cells

The notebook will:

1. Install dependencies
2. Load the base model
3. Prepare the dataset
4. Fine-tune (3 epochs, ~18 minutes)
5. Evaluate and create demo

📊 Training Results

After 3 epochs of training:

Epoch	Training Loss	Validation Loss	Token Accuracy
1	2.17	2.24	58.8%
2	1.25	2.28	58.9%
3	1.07	2.46	58.6%

🔑 Key Concepts

Full Fine-Tuning vs LORA

• Full Fine-Tuning: All model weights updated (used here)
• LORA: Only adapter weights trained (less resources needed)

SLM (Small Language Model)

• Models under 1B parameters
• Great for specific tasks
• Can be tailored for your use case

Tokens In, Tokens Out

• Think of any problem as: What tokens do I want in, and what tokens do I want out?

📁 Project Structure

├── NVIDIA-DGX-Spark-hugging_face_llm_full_fine_tune_tutorial-VIDEO.ipynb
├── README.md
├── .gitignore
└── Fully Fine-Tune a SLM-Gemma 3 270M.pdf (reference)

🎥 Video Tutorial

Watch the full video walkthrough

📚 Resources

• Original Notebook Source
• Hugging Face Transformers
• TRL Documentation
• Gemma 3 270M Model
• FoodExtract-1k Dataset

🏷️ Tags Dictionary

The model assigns these tags to text:

Tag	Meaning
np	Nutrition Panel
il	Ingredient List
me	Menu
re	Recipe
fi	Food Items
di	Drink Items
fa	Food Advertisement
fp	Food Packaging

📝 License

This project is for educational purposes. Please refer to the original sources for licensing information.

👤 Author

Created following the tutorial by Daniel Bourke

---

Note: The model weights are not included in this repository. You can either:

1. Run the notebook to create your own fine-tuned model
2. Download from Hugging Face Hub if available