---
language:
- vi
library_name: transformers
pipeline_tag: text-classification
base_model: "uitnlp/CafeBERT"
datasets:
- uitnlp/vihsd
tags:
- vietnamese
- text-classification
- vihsd
- transfer
- eacl-2027
- hate-speech-detection
- offensive-language
- social-media
metrics:
- f1
- accuracy
---

# cafebert-ViHSD

This model is `uitnlp/CafeBERT` fine-tuned for **ViHSD hate speech detection** on ViHSD.

## Evaluation protocol

- Dataset size: 33,400 examples.
- Original published fixed splits: 24,048 train / 2,672 development / 6,680 test.
- No rows were moved between the published splits.
- Fine-tuning seeds: [22, 42, 202].
- Training: 3 epoch(s), AdamW.
- Learning rate: 2e-05.
- Weight decay: 0.01.
- Warmup ratio: 0.1.
- Training batch size: 8.
- Maximum sequence length: 256.
- Input mode: raw Vietnamese social-media text.
- The published checkpoint is seed **22**.

## Results

Metrics are reported as mean ± sample standard deviation over the available completed seeds.

| Metric | Mean ± std |
|---|---:|
| Test Macro-F1 | 0.6430 ± 0.0197 |
| Test accuracy | 0.8739 ± 0.0064 |
| Test macro precision | 0.6718 ± 0.0203 |
| Test macro recall | 0.6312 ± 0.0160 |
| Development Macro-F1 | 0.6530 ± 0.0118 |

### Per-seed results

|       seed |   dev_macro_f1 |   test_macro_f1 |   test_accuracy |
|-----------:|---------------:|----------------:|----------------:|
|  22.000000 |       0.666226 |        0.656446 |        0.877246 |
|  42.000000 |       0.649098 |        0.620422 |        0.866467 |
| 202.000000 |       0.643704 |        0.652192 |        0.877994 |

## Label mapping

```json
{
  "0": "CLEAN",
  "1": "OFFENSIVE",
  "2": "HATE"
}
```

## Usage

```python
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "BaoNhan/cafebert-ViHSD"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "Đây là nội dung tiếng Việt cần phân loại."

inputs = tokenizer(
    text,
    return_tensors="pt",
    truncation=True,
    max_length=256,
)

with torch.no_grad():
    probabilities = model(**inputs).logits.softmax(dim=-1)[0]

predicted_id = int(probabilities.argmax())
print(model.config.id2label[predicted_id], probabilities.tolist())
```

## Limitations

ViHSD is class-imbalanced and reflects Vietnamese social-media language from a
particular collection period. Performance may not transfer directly to new
platforms, dialects, code-switching patterns, irony, or emerging slang.
Predictions should not be the sole basis for moderation or punitive decisions.

## Dataset citation

```bibtex
@InProceedings{10.1007/978-3-030-79457-6_35,
  author={Luu, Son T. and Nguyen, Kiet Van and Nguyen, Ngan Luu-Thuy},
  title={A Large-Scale Dataset for Hate Speech Detection on Vietnamese Social Media Texts},
  booktitle={Advances and Trends in Artificial Intelligence. Artificial Intelligence Practices},
  year={2021},
  publisher={Springer International Publishing},
  pages={415--426},
  doi={10.1007/978-3-030-79457-6_35}
}
```