hgthaianh/ViCoHate-Adaptive-Ultimate

ViCoHATE-Adaptive: Vietnamese Hate Speech Detection with Domain Adaptation

ViCoHATE-Adaptive is a robust Vietnamese Hate Speech Detection model designed to handle cross-domain challenges (Reddit to Facebook). It leverages Generative Augmentation, Pseudo-Labeling, and Simulated Active Learning to achieve state-of-the-art performance with minimal manual labeling.

This model is fine-tuned from FacebookAI/xlm-roberta-base using the ViCoHATE dataset and a proposed adaptive framework.

Model Details

Model Description

• Developed by: Hoàng Thái Anh
• Model type: Transformers (XLM-RoBERTa)
• Language(s): Vietnamese
• License: Apache 2.0
• Finetuned from model: FacebookAI/xlm-roberta-base
• Framework: PyTorch, Transformers

Key Features

1. Context-Aware: Inputs include Stance and Context (Post/Parent Comment) to resolve ambiguity.
2. Cross-Domain Robustness: Trained on a hybrid dataset of Reddit (Source), Augmented Facebook Data (Qwen3-generated), and Real Facebook Data (Pseudo-labeled).
3. Active Learning Refinement: Weighted training on difficult samples (Manual 1.6k) to boost Hate Recall.
4. Adaptive Thresholding: Optimized inference logic (Threshold = 0.15 for Hate class) to balance Safety and User Experience.

Uses

Direct Use

You can use this model directly for Hate Speech Detection in Vietnamese social media comments. It classifies text into three labels:

• 0: Clean (Nội dung sạch/Bình thường)
• 1: Offensive (Xúc phạm/Thô tục nhưng chưa đến mức thù ghét)
• 2: Hate (Thù ghét/Phân biệt chủng tộc/Chính trị cực đoan)

How to Get Started with the Model

1. Installation

pip install transformers torch

Inference Code (Optimized)

For the best performance, use the Adaptive Threshold logic below instead of standard argmax. This ensures that Hate Speech is detected with higher sensitivity (Recall) while maintaining safety.

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# Load Model
model_name = "hgthaianh/ViCoHate-Adaptive-Ultimate" 
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name).to("cuda" if torch.cuda.is_available() else "cpu")

def predict_optimized(text, context="", threshold=0.15):
    """
    Predicts label using Adaptive Thresholding for Hate class.
    Args:
        text (str): The comment to classify.
        context (str): Post content or parent comment (optional).
        threshold (float): Decision boundary for Hate class (Default: 0.15).
    """
    # Prepare Input
    full_text = f"Context: {context} | {text}" if context else text
    inputs = tokenizer(full_text, return_tensors="pt", truncation=True, max_length=256).to(model.device)
    
    # Inference
    with torch.no_grad():
        outputs = model(**inputs)
        probs = F.softmax(outputs.logits, dim=-1)[0].cpu().numpy()
    
    # Adaptive Logic
    # 0: Clean, 1: Offensive, 2: Hate
    if probs[2] >= threshold:
        label = "HATE"
    elif probs[0] > probs[1]:
        label = "CLEAN"
    else:
        label = "OFFENSIVE"
        
    return label, probs

# Test Example
text = "Lũ bắc kỳ chó này sống bẩn tính lắm"
context = "Bài viết về văn hóa vùng miền"
label, probs = predict_optimized(text, context)

print(f"Input: {text}")
print(f"Label: {label}")
print(f"Probabilities: Clean: {probs[0]:.2f}, Offensive: {probs[1]:.2f}, Hate: {probs[2]:.2f}")

Evaluation

Results (Test Set)

The model achieves a State-of-the-Art balance between Precision and Recall for the Hate class in a challenging cross-domain setting.

Metric	Score
Accuracy	71%
Macro F1	69%
Hate Recall	62%
Hate Precision	62%
Clean F1	79%

Citation

If you use this model or dataset, please cite our work:

@article{vicohate2026,
  title={},
  author={Hoang Thai Anh},
  journal={},
  year={2026}
}