🧠 Fine-tuning Generative LLMs to Detect and Explain Suicidal Ideation in Brazilian Portuguese Texts

This repository provides a complete pipeline for fine-tuning LLMs (Large Language Models) to classify suicidal ideation using 5-fold cross-validation, LoRA optimization, and ROC curve analysis. The models are trained with Unsloth for efficient memory usage and fine-tuning.

📌 Overview

This pipeline:

✔️ Uses Low-Rank Adaptation (LoRA) + RSLora for memory-efficient fine-tuning.
✔️ Trains and evaluates multiple LLMs, including LLaMA 3, Gemma, and Phi-3.
✔️ Implements 5-fold Stratified Cross-Validation for better model generalization.
✔️ Optimizes the classification threshold using the ROC-AUC curve.
✔️ Generates confusion matrices and performance reports automatically.

🚀 Technologies Used

The project relies on the following frameworks:

• Transformers 🤗 - NLP models and tokenizers
• TRL 🚀 - Fine-tuning with Reinforcement Learning
• Unsloth ⚡ - Optimized fine-tuning framework for LLaMA
• PEFT 🛠️ - Efficient parameter fine-tuning (LoRA, QLoRA)
• scikit-learn 📊 - Metrics, stratified cross-validation
• Matplotlib & Seaborn 📉 - Model visualization (ROC curve, confusion matrix)
• PyTorch 🔥 - Deep learning framework
• Datasets 📚 - Hugging Face dataset management
• Pandas & NumPy 🏗️ - Data manipulation

---

📂 Project Structure

1️⃣ Model Setup & Fine-Tuning

• Loads & configures LLMs (LLaMA 3, Phi-3, Gemma, Qwen)
• Modifies lm_head weights for classification tuning
• Uses LoRA with RSLora to efficiently fine-tune models

2️⃣ Data Processing

• Loads labeled dataset from CSV
• Splits dataset into training & validation (Stratified Sampling)
• Visualizes data distribution using Matplotlib

3️⃣ Training with Stratified Cross-Validation

• Uses 5-fold Stratified K-Fold
• Fine-tunes the model separately for each fold

4️⃣ ROC Curve & Optimal Threshold Calculation

• Generates ROC curve & AUC score
• Finds the optimal classification threshold
• Saves metrics & confusion matrix

5️⃣ Model Evaluation & Visualization

• Computes Accuracy, Precision, Recall, F1-score, AUC
• Saves & plots Confusion Matrix
• Plots ROC Curve to assess model performance

---

🛠 Installation & Setup

1️⃣ Install Required Dependencies

Make sure you have Python installed, then run:

pip install torch transformers trl unsloth datasets peft pandas numpy scikit-learn matplotlib seaborn tqdm