File size: 13,007 Bytes

---
license: llama2
datasets:
- yentinglin/zh_TW_c4
- yentinglin/traditional_mandarin_instructions
language:
- zh
widget:
 - text: "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: 你好，請問你可以幫我寫一封推薦信嗎？ ASSISTANT:"
library_name: transformers
pipeline_tag: text-generation
---
# Language Models for Taiwanese Culture 


<p align="center">
✍️ <a href="https://huggingface.co/spaces/yentinglin/Taiwan-LLaMa2" target="_blank">Online Demo</a>  
•
🤗 <a href="https://huggingface.co/yentinglin" target="_blank">HF Repo</a> • 🐦 <a href="https://twitter.com/yentinglin56" target="_blank">Twitter</a> • 📃 <a href="https://arxiv.org/pdf/2305.13711.pdf" target="_blank">[Paper Coming Soon]</a>  
• 👨️ <a href="https://yentingl.com/" target="_blank">Yen-Ting Lin</a> 
    <br/><br/>
    <img src="https://www.csie.ntu.edu.tw/~miulab/taiwan-llama/logo-v2.png" width="100"> <br/>
    <a href="https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE">
<img src="https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg"></a>
    <a href="https://github.com/tatsu-lab/stanford_alpaca/blob/main/DATA_LICENSE">
        <img src="https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-red.svg"></a>
    <br/>
   
</p>




## Overview
Taiwan-LLaMa is a full parameter fine-tuned model based on LLaMa 2 for Traditional Mandarin applications.

**Taiwan-LLaMa v1.0** pretrained on over 5 billion tokens and instruction-tuned on over 490k conversations both in traditional mandarin.

## Demo
A live demonstration of the model can be accessed at [Hugging Face Spaces](https://huggingface.co/spaces/yentinglin/Taiwan-LLaMa2).

## Key Features

1. **Traditional Mandarin Support**: The model is fine-tuned to understand and generate text in Traditional Mandarin, making it suitable for Taiwanese culture and related applications.

2. **Instruction-Tuned**: Further fine-tuned on conversational data to offer context-aware and instruction-following responses.

3. **Performance on Vicuna Benchmark**: Taiwan-LLaMa's relative performance on Vicuna Benchmark is measured against models like GPT-4 and ChatGPT. It's particularly optimized for Taiwanese culture.

4. **Flexible Customization**: Advanced options for controlling the model's behavior like system prompt, temperature, top-p, and top-k are available in the demo.


## Work in progress
- [ ] **Improved pretraining**: A refined pretraining process (e.g. more data from Taiwan, training strategies) is under development, aiming to enhance model performance for better Taiwanese culture.
- [ ] **Extend max length**: Utilizing the Rope mechanism as described in [the paper](https://arxiv.org/abs/2104.09864), the model's length will be extended from 4k to 8k.


## Taiwanese Culture Examples

| Question  | ChatGPT  | **Taiwan LLaMa v1.0**  | **Taiwan LLaMa v0.0 (w/o pretraining)**  | Llama2_Traditional_Chinese_13b_Chat  | 
|:-|:-|:-|:-|:-|
| 以惡地景觀著名的是高雄哪一個景點?   | 以惡地地景著名的是高雄市的旗津。    | 高雄以擁有惡地景觀著名的景點是田寮月世界。✅ | 以惡地景觀著名的是高雄市的旗津島。旗津島是一個小島，... | 高雄市內有多個以惡地景觀著名的景點，其中以下面幾個最為著名： 1. 高雄捷運紅線沿途景點：高雄市政府為了推廣高雄  |
| 唐鳳是誰？  | 唐鳳（Audrey Tang）是台灣的政治家、社會活動家和軟體工程師。她於2016年擔任台灣行政院數位政務委員，是台灣政府歷史上首位沒有政黨背景的政務官。唐鳳在軟體工程領域有豐富的經驗。 ✅ | 唐鳳是政治評論家、資訊社會學研究者，也是台灣第一位擔任政府數位政委的人。 ✅ | 唐鳳（Tang Feng）是一位著名的中國演員和歌手。他出生於1974年12月17日，出生地點在中國北京。唐鳳在1995年開始他的演藝生涯，並在同年發行了他的首張專輯《我不孤單》。此後，他發行了許多受歡迎的歌曲，並在多部電視劇和電影中擔任主角。唐鳳因其精湛的演技和迷人的歌聲而被譽為“亞洲魅力男王”。他的音樂獲得了多個獎項，包括“華表最佳新男演唱獎”、“華表最佳歌曲獎”和“華表最佳舞蹈獎”等。  | 唐鳳是唐朝時期的一位著名文學家，他的代表作品包括《陰騭》、《狂凡子》等。他的文學風格具有強烈的個性和創意，被譽為「唐代文學之父」。  |
| NTU 在哪?                      | NTU位於新加坡，具體地址是50 Nanyang Avenue, Singapore 639798。   | NTU 是國立臺灣大學的縮寫。它是臺灣的一所頂尖研究型大學，位於臺北市。 ✅| NTU 是台灣的一所大學，位於台北市。✅ | NTU 的總部位於台灣新北市板橋區，佔地面積約為1,000公頃。|
| 甚麼是22K? | "22K" 可能指的是金的純度，尤其是在珠寶和金飾方面。 | "22K"是指台灣的大學畢業生的起薪，大約是新台幣22,000元。這個詞被用來形容大學畢業生的起薪偏低，並且在就業市場上對於大學畢業生的需求不足。✅| 22K是指溫度尺度上的22,000K（千度）。1K代表一度，即溫度差距，是攝氏溫標（攝氏度）的一百萬分之一。因此，22K等於-253.15°C（攝氏溫標）或-423.67°F（華氏溫標）。 |  22K是一種簡寫，即22,000。它是一個數字，表示22,000個單位。|
| 肉圓要不要加香菜？                  | 這是個個人口味的問題，有些人喜歡在肉圓中加入香菜，因為香菜的清香可以增添風味；而有些人則不喜歡香菜的味道，覺得它會蓋過其他食材的味道。所以要不要加香菜，可以根據個人喜好來決定。 | 這是一個個人喜好的問題，有些人喜歡在肉圓裡加香菜，有些人不喜歡。**在台灣**，肉圓加香菜是一種常見的做法，這是一種在肉圓上加香菜的做法。然而，在其他國這是一個文化差異的問題。😊 | 肉圓加香菜可以增添一些芳香和美味的口感，如果你喜歡香菜的味道，可以試試加入肉圓中。香菜通常在製作過程中炒香，可以增添一些芳香和美味的味道。如果你想要加入香菜，可以在製作過程中炒香一些香菜，然後將其混入肉圓中。肉圓加香菜可以增添一些芳香和美味的口感，如果你喜歡香菜的味道，可以試試加入肉圓中。| 加香菜的肉圓味道更好，但是不加香菜的肉圓也很好吃。如果你喜歡香菜的味道，那麼加上香菜就好了。如果你不喜歡香菜，那麼不加也不會影響肉 |


## Model

We provide a number of model checkpoints that we trained. Please find them on Hugging Face [here](https://huggingface.co/models?search=taiwan-llama). Here are some quick links to the checkpoints that are finetuned from LLaMa 2:

| **Model**                                              | **13B**                                                                                                                       | 
|--------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------|
| **Taiwan-LLaMa v1.0** (_better for Taiwanese Culture_)   | 🤗 <a href="https://huggingface.co/yentinglin/Taiwan-LLaMa-v1.0" target="_blank">yentinglin/Taiwan-LLaMa-v1.0</a>  | 
| Taiwan-LLaMa v0.9 (partial instruction set)            | 🤗 <a href="https://huggingface.co/yentinglin/Taiwan-LLaMa-v0.9" target="_blank">yentinglin/Taiwan-LLaMa-v0.9</a>  | 
| Taiwan-LLaMa v0.0 (no Traditional Mandarin pretraining) | 🤗 <a href="https://huggingface.co/yentinglin/Taiwan-LLaMa-v0.0" target="_blank">yentinglin/Taiwan-LLaMa-v0.0</a>  | 

## Data

Here are some quick links to the datasets that we used to train the models:

| **Dataset**                     | **Link**                                                                                                                      | 
|---------------------------------|-------------------------------------------------------------------------------------------------------------------------------|
| **Instruction-tuning**      | 🤗 <a href="https://huggingface.co/datasets/yentinglin/traditional_mandarin_instructions" target="_blank">yentinglin/traditional_mandarin_instructions</a>                                           | 
| Traditional Mandarin Pretraining | 🤗 <a href="https://huggingface.co/datasets/yentinglin/zh_TW_c4" target="_blank">yentinglin/zh_TW_c4</a>                                   | 


## Architecture
Taiwan-LLaMa is based on LLaMa 2, leveraging transformer architecture, <a href="https://github.com/Dao-AILab/flash-attention" target="_blank">flash attention 2</a>, and bfloat16.

It includes:

* Pretraining Phase: Pretrained on a vast corpus of over 5 billion tokens, extracted from common crawl in Traditional Mandarin.
* Fine-tuning Phase: Further instruction-tuned on over 490k multi-turn conversational data to enable more instruction-following and context-aware responses.

## Generic Capabilities on Vicuna Benchmark 

The data is translated into traditional mandarin for evaluating the general capability.


<img src="./images/zhtw_vicuna_bench_chatgptbaseline.png" width="700">

The scores are calculated with ChatGPT as the baseline, represented as 100%. The other values show the relative performance of different models compared to ChatGPT.

| Language Model                      | Relative Score (%) |
|-------------------------------------|--------------------|
| GPT-4                               | 102.59%            |
| ChatGPT                             | 100.00%            |
| **Taiwan-LLaMa v1.0**               | 76.76%             |
| Claude-Instant-1.2                  | 74.04%             |
| Llama2_Traditional_Chinese_13b_Chat | 56.21%             |




## How to deploy the model on my own machine?
We recommend hosting models with [🤗 Text Generation Inference](https://github.com/huggingface/text-generation-inference). Please see their [license](https://github.com/huggingface/text-generation-inference/blob/main/LICENSE) for details on usage and limitations.
```bash
bash run_text_generation_inference.sh "yentinglin/Taiwan-LLaMa" NUM_GPUS DIR_TO_SAVE_MODEL PORT MAX_INPUT_LEN MODEL_MAX_LEN
```

Prompt format follows vicuna-v1.1 template:
```
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: {user} ASSISTANT:
```

## Setup development environment
```bash
conda create -n taiwan-llama python=3.10 -y 
conda activate taiwan-llama
pip install -r requirements.txt
```


## Citations
If you use our code, data, or models in your research, please cite this repository. You can use the following BibTeX entry:

```bibtex
@inproceedings{lin-chen-2023-llm,
    title = "{LLM}-Eval: Unified Multi-Dimensional Automatic Evaluation for Open-Domain Conversations with Large Language Models",
    author = "Lin, Yen-Ting  and Chen, Yun-Nung",
    booktitle = "Proceedings of the 5th Workshop on NLP for Conversational AI (NLP4ConvAI 2023)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.nlp4convai-1.5",
    pages = "47--58"
}

@misc{taiwanllama,
    author={Lin, Yen-Ting and Chen, Yun-Nung},
    title={Taiwanese-Aligned Language Models based on Meta-Llama2},
    year={2023},
    url={https://github.com/adamlin120/Taiwan-LLaMa},
    note={Code and models available at https://github.com/adamlin120/Taiwan-LLaMa},
}
```

## Collaborate With Us
If you are interested in contributing to the development of Traditional Mandarin language models, exploring new applications, or leveraging Taiwan-LLaMa for your specific needs, please don't hesitate to contact us. We welcome collaborations from academia, industry, and individual contributors.

## License
The code in this project is licensed under the Apache 2.0 License - see the [LICENSE](LICENSE) file for details.

The models included in this project are licensed under the LLAMA 2 Community License. See the [LLAMA2 License](https://github.com/facebookresearch/llama/blob/main/LICENSE) for full details.

## OpenAI Data Acknowledgment
The data included in this project were generated using OpenAI's models and are subject to OpenAI's Terms of Use. Please review [OpenAI's Terms of Use](https://openai.com/policies/terms-of-use) for details on usage and limitations.


## Acknowledgements

We thank [Meta LLaMA team](https://github.com/facebookresearch/llama) and [Vicuna team](https://github.com/lm-sys/FastChat) for their open-source efforts in democratizing large language models.