https://github.com/madhans476/food-hazard-detection-semeval-2025

https://github.com/madhans476/food-hazard-detection-semeval-2025

Last synced: 14 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/madhans476/food-hazard-detection-semeval-2025
• Owner: madhans476
• License: mit
• Created: 2025-02-14T06:57:18.000Z (4 months ago)
• Default Branch: main
• Last Pushed: 2025-02-26T06:41:26.000Z (3 months ago)
• Last Synced: 2025-02-26T07:31:50.570Z (3 months ago)
• Size: 2.99 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Food Hazard Detection - SemEval 2025 Task 9  

## 📌 Project Overview  

This repository contains our solution for **SemEval-2025 Task 9: The Food Hazard Detection Challenge**, which aims to **automate food hazard classification and extraction** from recall notifications.  

The project consists of **two subtasks**:  

1. **Task 1 (Classification)**: Predicts the **hazard category** (e.g., Biological, Chemical) and **product category** (e.g., Dairy, Meat).  

2. **Task 2 (Detection)**: Extracts the **exact hazard reason** and **product name** from recall text.  

We employ **machine learning, fine-tuned LLMs (GPT-2, LLaMA, Flan-T5-XL), and ensemble learning** to solve these challenges.  

## Project Structure

```plaintext

Food-Hazard-Detection-SemEval-2025/ 

│── Datasets/                     # Dataset storage (train, validation, test) 

│   │── train.csv                 # Preprocessed training data 

│   │── validation.csv            # Raw validation dataset 

│   │── test.csv                  # Raw test dataset 

│   │── README.md                 # Dataset details and preprocessing steps 

│── Task 1/                       # Task 1: Hazard & Product Category Classification 

│   │── TF-IDF_XGBoost.ipynb      # Train and evaluate TF-IDF + XGBoost model 

│   │── peft_GPT2.ipynb           # Fine-tune GPT-2 Large using PEFT 

│   │── peft_llama.ipynb          # Fine-tune LLaMA 3.1 1B using PEFT 

│   │── ensemble.py               # Combine models using a voting ensemble 

│   │── README.md                 # Task 1 details 

│── Task 2/                       # Task 2: Hazard & Product Extraction 

│   │── prompt_based_peft.ipynb   # Fine-tune Flan-T5-XL for extraction 

│   │── README.md                 # Task 2 details 

│── LICENSE                       # License information 

│── README.md                     # Main project documentation 

└── requirements.txt              # Python dependencies 

```

## 📂 Dataset Details  

- **Source**: The dataset contains food recall notifications scraped from regulatory sources.  

- **Language**: English.  

- **Size**: ~7,546 recall notices.  

- **Labels**:  

  - **Hazard Categories**: 10 classes (e.g., Biological, Chemical, Physical).  

  - **Product Categories**: 22 classes (e.g., Beverages, Dairy, Meat).

  - **product** → The exact food product name in the recall notice.  

  - **hazard** → The exact hazard reason in the recall notice.  

📌 **More details**: See [`Datasets/README.md`](Datasets/readme.md).  

## ⚙️ Methodology  

### **🛠 Task 1: Hazard & Product Category Classification**  

🔹 **Goal**: Assign a recall notice to a **hazard category** & **product category**.  

🔹 **Models Used**:  

✅ **TF-IDF + XGBoost** (Baseline).  

✅ **Fine-Tuned GPT-2 & LLaMA (LoRA-PEFT)**.  

✅ **Ensemble Learning (Voting Classifier)** for hazard classification.  

📌 **More details**: See [`Task 1/README.md`](https://github.com/madhans476/Food-hazard-detection-SEMEVAL-2025/blob/fda1f55214af673a48ef485a72c563ee9135e17d/Task%201/readme.md).  

### **🔍 Task 2: Hazard & Product Extraction**  

🔹 **Goal**: Extract the **exact hazard reason** and **product name** from recall text.  

🔹 **Model Used**:  

✅ **Flan-T5-XL** (Fine-tuned with PEFT & LoRA).  

✅ **Prompt Engineering** for improved extraction.  

📌 **More details**: See [`Task 2/README.md`](https://github.com/madhans476/Food-hazard-detection-SEMEVAL-2025/blob/fda1f55214af673a48ef485a72c563ee9135e17d/Task%202/readme.md).  

## 🚀 Installation & Setup  

### **1️⃣ Install Dependencies**  

Ensure you have **Python 3.8+** installed. Then, run:  

```bash

pip install -r requirements.txt

```

### 2️⃣ Pre-Run Checklist

- Ensure dataset file path.

- Set up Hugging Face API token (for loading models).

- Set up WandB API key (if using experiment tracking).

- Verify fine-tuned model paths before execution.

## Acknowledgment

We acknowledge the authors for providing the dataset:

**Dataset Citation:**

Randl, Korbinian; Karvounis, Manos; Marinos, George; Pavlopoulos, John; Lindgren, Tony; Henriksson, Aron.  

*Food Recall Incidents*.  

Publisher: Zenodo, March 2024.  

DOI: [10.5281/zenodo.10891602](https://doi.org/10.5281/zenodo.10891602)