visolex/bartpho-hsd

1---2license: mit3base_model: vinai/bartpho-syllable-base4tags:5- vietnamese6- hate-speech-detection7- text-classification8- offensive-language-detection9datasets:10- visolex/vihsd11metrics:12- accuracy13- macro-f114- weighted-f115model-index:16- name: bartpho-hsd17  results:18  - task:19      type: text-classification20      name: Hate Speech Detection21    dataset:22      name: ViHSD23      type: hate-speech-detection24    metrics:25    - type: accuracy26      value: 0.898527    - type: macro-f128      value: 0.679129    - type: weighted-f130      value: 0.888631    - type: macro-precision32      value: 0.766433    - type: macro-recall34      value: 0.628935---36 37# BARTpho: Hate Speech Detection for Vietnamese Text38 39This model is a fine-tuned version of [vinai/bartpho-syllable-base](https://huggingface.co/vinai/bartpho-syllable-base) 40on the **ViHSD (Vietnamese Hate Speech Detection Dataset)** for classifying Vietnamese text into three categories: CLEAN, OFFENSIVE, and HATE.41 42## Model Details43 44* **Base Model**: vinai/bartpho-syllable-base45* **Description**: BARTpho fine-tuned cho bài toán phân loại Hate Speech tiếng Việt46* **Architecture**: BARTpho (Bidirectional and Auto-Regressive Transformer cho tiếng Việt)47* **Dataset**: ViHSD (Vietnamese Hate Speech Detection Dataset)48* **Fine-tuning Framework**: HuggingFace Transformers + PyTorch49* **Task**: Hate Speech Classification (3 classes)50 51### Hyperparameters52 53* **Batch size**: `32`54* **Learning rate**: `2e-5`55* **Epochs**: `100`56* **Max sequence length**: `256`57* **Weight decay**: `0.01`58* **Warmup steps**: `500`59* **Early stopping patience**: `5`60* **Optimizer**: AdamW61* **Learning rate scheduler**: Cosine with warmup62 63## Dataset64 65Model was trained on **ViHSD (Vietnamese Hate Speech Detection Dataset)** containing ~10,000 Vietnamese comments from social media.66 67### Label Descriptions:68 69* **CLEAN (0)**: Normal content without offensive language70* **OFFENSIVE (1)**: Mildly offensive or inappropriate content  71* **HATE (2)**: Hate speech, extremist language, severe threats72 73## Evaluation Results74 75The model was evaluated on test set with the following metrics:76 77* **Accuracy**: `0.8985`78* **Macro-F1**: `0.6791`79* **Weighted-F1**: `0.8886`80* **Macro-Precision**: `0.7664`81* **Macro-Recall**: `0.6289`82 83### Basic Usage84 85```python86from transformers import AutoTokenizer, AutoModelForSequenceClassification87import torch88 89# Load model and tokenizer90model_name = "visolex/bartpho-hsd"91tokenizer = AutoTokenizer.from_pretrained(model_name)92model = AutoModelForSequenceClassification.from_pretrained(93    model_name94)95 96# Classify text97text = "Văn bản tiếng Việt cần phân loại"98inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)99 100with torch.no_grad():101    outputs = model(**inputs)102    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)103    predicted_label = torch.argmax(predictions, dim=-1).item()104 105# Label mapping106label_names = {107    0: "CLEAN",108    1: "OFFENSIVE",109    2: "HATE"110}111 112print(f"Predicted label: {label_names[predicted_label]}")113print(f"Confidence scores: {predictions[0].tolist()}")114```115 116 117 118## Training Details119 120### Training Data121- **Dataset**: ViHSD (Vietnamese Hate Speech Detection Dataset)122- **Total samples**: ~10,000 Vietnamese comments from social media123- **Training split**: ~70%124- **Validation split**: ~15%125- **Test split**: ~15%126 127### Training Configuration128- **Framework**: PyTorch + HuggingFace Transformers129- **Optimizer**: AdamW130- **Learning Rate**: 2e-5131- **Batch Size**: 32132- **Max Length**: 256 tokens133- **Epochs**: 100 (with early stopping patience: 5)134- **Weight Decay**: 0.01135- **Warmup Steps**: 500136 137 138## Contact & Support139 140- **GitHub**: [ViSoLex Hate Speech Detection](https://github.com/visolex/hate-speech-detection)141- **Issues**: [Report Issues](https://github.com/visolex/hate-speech-detection/issues)142- **Questions**: Open a discussion on the model's Hugging Face page143 144## License145 146This model is distributed under the MIT License.147 148## Acknowledgments149 150- Base model: [vinai/bartpho-syllable-base](https://huggingface.co/vinai/bartpho-syllable-base)151- Dataset: ViHSD (Vietnamese Hate Speech Detection Dataset)152- Framework: [Hugging Face Transformers](https://huggingface.co/transformers)153- ViSoLex Toolkit154 155---156