---
language: vi
library_name: transformers
pipeline_tag: text-classification
tags:
- clickbait-detection
- vietnamese
- viclickbait-2025
datasets:
- ViClickbait-2025
---

# CafeBERT — ViClickbait-2025

Fine-tuned from `uitnlp/CafeBERT` for binary Vietnamese clickbait detection.

## Experimental setup

- Input: headline paired with lead paragraph; no URL, source, category, publish time, image, or engagement metadata.
- Fixed 80/10/10 split using `StratifiedGroupKFold` with seed 42.
- Fine-tuning seeds: [42, 22, 202]; 3 epochs per seed.
- Development Macro-F1 selects checkpoints and representative seed.
- Weighted cross-entropy from training-label frequencies: `True`.
- Effective batch size: 8; max length: 256.


## Results

| Metric | Mean ± sample std |
|---|---:|
| Test Macro-F1 | 0.8047 ± 0.0060 |
| Test accuracy | 0.8236 ± 0.0074 |
| Dev Macro-F1 | 0.8256 ± 0.0053 |

Representative seed: **22**, selected only by development Macro-F1.

### Per-seed

|   seed |   dev_macro_f1 |   test_macro_f1 |   test_accuracy |
|-------:|---------------:|----------------:|----------------:|
|     22 |         0.8312 |          0.8043 |          0.8246 |
|     42 |         0.8249 |          0.7989 |          0.8158 |
|    202 |         0.8207 |          0.8108 |          0.8304 |

## Labels

- `0`: non-clickbait
- `1`: clickbait

## Usage

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "BaoNhan/cafebert-ViClickbait-2025"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
title = "Tiêu đề bài báo"
lead = "Đoạn dẫn của bài báo"
inputs = tokenizer(title, lead, return_tensors="pt", truncation=True, max_length=256)
prediction = model(**inputs).logits.argmax(dim=-1).item()
print(model.config.id2label[prediction])
```

## Dataset

- Nguyen et al. (2025), *ViClickbait-2025: A comprehensive dataset for Vietnamese clickbait detection*. https://doi.org/10.1016/j.dib.2025.112164
- Dataset: https://doi.org/10.17632/3wc46bfcjc.1

## Limitations

The dataset is small, temporally bounded to 2023–2025, and collected from eight Vietnamese news platforms. Results may not transfer to social media, other publishers, or emerging clickbait styles.