Skip to content

MODELS.md

Latest commit

 

History

History
115 lines (87 loc) · 3.51 KB

MODELS.md

File metadata and controls

115 lines (87 loc) · 3.51 KB
Raw

Model Files and Download Instructions

📦 Model Files

Due to GitHub's file size limitations (100MB per file), the trained models are stored separately.

Model Information

Language Model File Size Non-O F1 Download
Chinese chinese_xlmr_model.pt ~1.1GB 69.89% Download
English english_xlmr_model.pt ~1.1GB 73.65% Download
Korean korean_xlmr_model.pt ~1.1GB 67.56% Download

🚀 Quick Setup

Option 1: Download Models Manually

# Create models directory
mkdir -p backend/models/

# Download models (replace with actual URLs)
wget https://example.com/chinese_xlmr_model.pt -O backend/models/chinese_xlmr_model.pt
wget https://example.com/english_xlmr_model.pt -O backend/models/english_xlmr_model.pt
wget https://example.com/korean_xlmr_model.pt -O backend/models/korean_xlmr_model.pt

Option 2: Train Models Yourself

# Train all models (will take several hours)
python backend/training/chinese_xlmr_train.py --epochs 8 --patience 3
python backend/training/english_xlmr_train.py --epochs 8 --patience 3
python backend/training/korean_xlmr_train.py --epochs 8 --patience 3

Option 3: Use Hugging Face Hub (Recommended)

from transformers import AutoModel
import torch

# Load model from Hugging Face Hub
model = AutoModel.from_pretrained("USERNAME/chinese-xlmr-hate-detection")

🔧 Model Usage

Loading Trained Models

import torch
from backend.training.chinese_xlmr_train import ChineseXLMDetector

# Load model
model = ChineseXLMDetector(num_targets=6)
model.load_state_dict(torch.load('backend/models/chinese_xlmr_model.pt'))
model.eval()

# Use for inference
# (See individual training scripts for usage examples)

Model Specifications

Chinese Model

  • Base: XLM-RoBERTa-base
  • Targets: 6 categories (Sexism, Racism, LGBTQ, Region, others, non-hate)
  • BIO Classes: 5 (O, B-MILD, I-MILD, B-SEVERE, I-SEVERE)
  • Intensity: 3 levels (NONE, MILD, SEVERE)

English Model

  • Base: XLM-RoBERTa-base
  • Targets: 10 categories
  • BIO Classes: 9 (O, B-SOFT, I-SOFT, B-HARD, I-HARD, B-MILD, I-MILD, B-SEVERE, I-SEVERE)
  • Intensity: 6 levels (normal, offensive, L1_hate, L2_hate, mild, severe)

Korean Model

  • Base: XLM-RoBERTa-base
  • Targets: Not available in current implementation
  • BIO Classes: 5 (O, B-MILD, I-MILD, B-SEVERE, I-SEVERE)
  • Intensity: 4 levels

📊 Performance Benchmarks

Training Configuration

  • Optimizer: AdamW (lr=2e-5)
  • Batch Size: 16 (train), 32 (test)
  • Max Epochs: 8
  • Early Stopping: Patience 3
  • Class Weights: Applied for BIO loss

Results Summary

  • Best Non-O F1: English (73.65%)
  • Best BIO F1: Korean (97.27%)
  • Best Intensity F1: Chinese (78.69%)
  • Most Stable: Korean (largest dataset)

🔄 Model Updates

Models are periodically updated with:

  • Better hyperparameters
  • Improved preprocessing
  • Additional training data
  • Architecture optimizations

Check the release notes for update information.

📞 Support

For model-related issues:

  • Check the training logs in backend/results/
  • Review the training scripts for configuration
  • Open an issue with model performance questions

Note: Models are trained on specific datasets and may not generalize to all domains. Consider fine-tuning for your specific use case.