---
base_model: tarudesu/ViHateT5-base
tags:
- generated_from_trainer
model-index:
- name: ViHateT5-base-HSD
  results: []
datasets:
- tarudesu/ViCTSD
- tarudesu/ViHOS
- tarudesu/ViHSD
language:
- vi
metrics:
- f1
- accuracy
pipeline_tag: text2text-generation
widget:
- text: "toxic-speech-detection: Nhìn bà không thể không nhớ đến các phim phù thủy"
- text: "hate-speech-detection: thằng đó trông đần vcl ấy nhỉ"
- text: "hate-spans-detection: trông như cl"
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# <a name="introduction"></a>ViHateT5: Enhancing Hate Speech Detection in Vietnamese with A Unified Text-to-Text Transformer Model | ACL'2024 (Findings)
**Disclaimer**: This paper contains examples from actual content on social media platforms that could be considered toxic and offensive.

ViHateT5-HSD is the fine-tuned model of [ViHateT5](https://huggingface.co/tarudesu/ViHateT5-base) on multiple Vietnamese hate speech detection benchmark datasets.

The architecture and experimental results of ViHateT5 can be found in the [paper](LINK):

    @misc{nguyen2024vihatet5,
          title={ViHateT5: Enhancing Hate Speech Detection in Vietnamese With A Unified Text-to-Text Transformer Model}, 
          author={Luan Thanh Nguyen},
          year={2024},
          eprint={2405.14141},
          archivePrefix={arXiv},
          primaryClass={cs.CL}
    }
    
The pre-training dataset named VOZ-HSD is available at [HERE](https://huggingface.co/datasets/tarudesu/VOZ-HSD).

Kindly **CITE** our paper if you use ViHateT5-HSD to generate published results or integrate it into other software.

**Example usage**
```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("tarudesu/ViHateT5-base-HSD")
model = AutoModelForSeq2SeqLM.from_pretrained("tarudesu/ViHateT5-base-HSD")

def generate_output(input_text, prefix):
    # Add prefix
    prefixed_input_text = prefix + ': ' + input_text

    # Tokenize input text
    input_ids = tokenizer.encode(prefixed_input_text, return_tensors="pt")

    # Generate output
    output_ids = model.generate(input_ids, max_length=256)

    # Decode the generated output
    output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)

    return output_text

sample = 'Tôi ghét bạn vl luôn!'
prefix = 'hate-spans-detection' # Choose 1 from 3 prefixes ['hate-speech-detection', 'toxic-speech-detection', 'hate-spans-detection']

result = generate_output(sample, prefix)
print('Result: ', result)
```

Please feel free to contact us by email luannt@uit.edu.vn if you have any further information!