README.md

---
license: cc-by-nc-sa-4.0
---
<p align="center">
<img src="https://github.com/lbh0830/TW-Hokkien-LLM/blob/main/pics/logo.jpg?raw=true" alt="Taigi-llama-logo" width="350">
</p>

# Model Card for Taigi-Llama-2-13B
The Taigi-Llama-2 series are built based on the Traditional Chinese version of the LLaMA-2 model. We conducted continued pre-training on web-scraped data in Taiwanese Hokkien, including Hanzi, POJ, and Hanlo, totaling around 78MB.

For more details, please refer to our [GitHub repository](https://github.com/lbh0830/TW-Hokkien-LLM/tree/main) and the paper: [Enhancing Taiwanese Hokkien Dual Translation by Exploring and Standardizing of Four Writing Systems](https://arxiv.org/abs/2403.12024)

Explore other models and datasets in the [Taiwanese Hokkien LLM collection](https://huggingface.co/collections/Bohanlu/taiwanese-hokkien-llm-6614ba7456e6789bc2f10ca0).

## Model description

- **Usage:** This model can be used for causal language modeling tasks in Taiwanese Hokkien. It is also suitable for further fine-tuning on specific datasets for downstream tasks.
- **Language(s) (NLP):** The primary language is Taiwanese Hokkien (Hanzi and POJ). The model also retains capabilities in English and Mandarin Chinese due to prior pre-training.
- **Input:** Text
- **Output:** Text
- **Model Size:** 13B parameters

## Usage Example
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TextGenerationPipeline
import torch
import accelerate

def get_pipeline(path:str, tokenizer:AutoTokenizer, accelerator:accelerate.Accelerator) -> TextGenerationPipeline:
    model = AutoModelForCausalLM.from_pretrained(
        path, torch_dtype=torch.float16, device_map='auto', trust_remote_code=True)
    
    terminators = [tokenizer.eos_token_id, tokenizer.pad_token_id]

    pipeline = TextGenerationPipeline(model = model, tokenizer = tokenizer, num_workers=accelerator.state.num_processes*4, pad_token_id=tokenizer.pad_token_id, eos_token_id=terminators)

    return pipeline

model_dir = "Bohanlu/Taigi-Llama-2-7B" # or Bohanlu/Taigi-Llama-2-13B for the 13B model
tokenizer = AutoTokenizer.from_pretrained(model_dir, use_fast=False)

accelerator = accelerate.Accelerator()
pipe = get_pipeline(model_dir, tokenizer, accelerator)

# Few-shot示例：問答
qa_prompt = """Example 1:
問題：台北101有偌懸？
答案：台北101的高度是五百空八公尺。

Example 2:
問題：台灣上長的溪仔是佗一條？
答案：台灣上長的溪仔是濁水溪，規个長度有百八公里遐爾長。

Example 3:
問題：臺灣上懸的山是啥物？
答案："""

print(pipe(qa_prompt, return_full_text=False))
# Output: [{'generated_text': '臺灣上懸的山是玉山，海拔三千九百五十二公尺。'}]
```

## Citation

If you find the resources in the Taiwanese Hokkien LLM collection useful in your work, please cite it using the following reference:

```
@misc{lu2024enhancing,
      title={Enhancing Taiwanese Hokkien Dual Translation by Exploring and Standardizing of Four Writing Systems}, 
      author={Bo-Han Lu and Yi-Hsuan Lin and En-Shiun Annie Lee and Richard Tzong-Han Tsai},
      year={2024},
      eprint={2403.12024},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```