CurupiraIA: Hate Speech Detection ML Model 🇧🇷🛡️

Table of Contents

• Project Description
• Current Features
• Installation
• Dataset
• Model Architecture
• Training
• Contributors
• License

Project Description 🔍

Portuguese-language hate speech detection model using BERT architecture, currently in development phase. Key aspects:

• Fine-tuning neuralmind/bert-base-portuguese-cased
• Binary classification (hate speech detection)
• Focus on Brazilian social media text patterns
• Experimental phase with HateBR dataset

Current Features ✅

• Data preprocessing pipeline
• BERT model fine-tuning setup
• Basic evaluation metrics
• Hugging Face integration
• Experiment tracking (W&B)

Installation ⚙️

Prerequisites

• Python 3.10+
• CUDA-enabled GPU (recommended)

git clone https://github.com/yourusername/curupira-ml.git
cd curupira-ml
pip install -r requirements.txt

Dataset 📊

HateBR Dataset

from datasets import load_dataset

dataset = load_dataset("ruanchaves/hatebr")

Source: Hugging Face Datasets
7,000 annotated comments
Features:
Text: Raw comment text
Label: Binary classification (0=normal, 1=hate)

Model Architecture 🧠

Base Configuration

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=2)

Fine-Tuning Phase 🎯

Current Status:

[2025] Model in active fine-tuning stage

We're currently optimizing the BERT-large model through:

• Progressive learning rate decay
• Class imbalance mitigation techniques
• Early stopping implementation
• Validation metric monitoring (F1-score focus)

Training 🔥

Traning Command

trainer.train()

Contributors

Kim Gomes

Licença

Copyright ©️ 2025 - Curupira AI