yentinglin commited on
Commit
bbcc1a7
1 Parent(s): e7733ed

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +164 -1
README.md CHANGED
@@ -7,4 +7,167 @@ language:
7
  - zh
8
  library_name: transformers
9
  pipeline_tag: text-generation
10
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7
  - zh
8
  library_name: transformers
9
  pipeline_tag: text-generation
10
+ ---
11
+ # Language Models for Taiwanese Culture
12
+
13
+
14
+ <p align="center">
15
+ ✍️ <a href="https://huggingface.co/spaces/yentinglin/Taiwan-LLaMa2" target="_blank">Online Demo</a>
16
+
17
+ 🤗 <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>
18
+ • 👨️ <a href="https://yentingl.com/" target="_blank">Yen-Ting Lin</a>
19
+ <br/><br/>
20
+ <img src="https://www.csie.ntu.edu.tw/~miulab/taiwan-llama/logo-v2.png" width="100"> <br/>
21
+ <a href="https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE">
22
+ <img src="https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg"></a>
23
+ <a href="https://github.com/tatsu-lab/stanford_alpaca/blob/main/DATA_LICENSE">
24
+ <img src="https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-red.svg"></a>
25
+ <br/>
26
+
27
+ </p>
28
+
29
+
30
+
31
+
32
+ ## Overview
33
+ Taiwan-LLaMa is a full parameter fine-tuned model based on LLaMa 2 for traditional chinese applications.
34
+
35
+ **Taiwan-LLaMa v1.0** pretrained on over 5 billion tokens and instruction-tuned on over 490k conversations both in traditional chinese.
36
+
37
+ ## Demo
38
+ A live demonstration of the model can be accessed at [Hugging Face Spaces](https://huggingface.co/spaces/yentinglin/Taiwan-LLaMa2).
39
+
40
+ ## Key Features
41
+
42
+ 1. **Traditional Chinese Support**: The model is fine-tuned to understand and generate text in Traditional Chinese, making it suitable for Taiwanese culture and related applications.
43
+
44
+ 2. **Instruction-Tuned**: Further fine-tuned on conversational data to offer context-aware and instruction-following responses.
45
+
46
+ 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.
47
+
48
+ 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.
49
+
50
+
51
+ ## Work in progress
52
+ - [ ] **Improved Pretraining**: A refined version of the existing pretraining approach is under development, aiming to enhance model performance.
53
+ - [ ] **Extended Model 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.
54
+
55
+
56
+ ## Taiwanese Culture Examples
57
+
58
+ | Question | ChatGPT | **Taiwan LLaMa v1.0** | **Taiwan LLaMa v0.0 (w/o pretraining)** | Llama2_Traditional_Chinese_13b_Chat |
59
+ |:-|:-|:-|:-|:-|
60
+ | 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 |
61
+ | 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」。 |
62
+ | 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。|
63
+ | 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���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。|
64
+ | 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 |
65
+
66
+
67
+ ## Model
68
+
69
+ 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:
70
+
71
+ | **Model** | **13B** |
72
+ |--------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------|
73
+ | **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> |
74
+ | 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> |
75
+ | Taiwan-LLaMa v0.0 (no Traditional Chinese pretraining) | 🤗 <a href="https://huggingface.co/yentinglin/Taiwan-LLaMa-v0.0" target="_blank">yentinglin/Taiwan-LLaMa-v0.0</a> |
76
+
77
+ ## Data
78
+
79
+ Here are some quick links to the datasets that we used to train the models:
80
+
81
+ | **Dataset** | **Link** |
82
+ |---------------------------------|-------------------------------------------------------------------------------------------------------------------------------|
83
+ | **Instruction-tuning** | 🤗 <a href="https://huggingface.co/datasets/yentinglin/traditional_chinese_instructions" target="_blank">yentinglin/traditional_chinese_instructions</a> |
84
+ | Traditional Chinese Pretraining | 🤗 <a href="https://huggingface.co/datasets/yentinglin/zh_TW_c4" target="_blank">yentinglin/zh_TW_c4</a> |
85
+
86
+
87
+ ## Architecture
88
+ 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.
89
+
90
+ It includes:
91
+
92
+ * Pretraining Phase: Pretrained on a vast corpus of over 5 billion tokens, extracted from common crawl in Traditional Chinese.
93
+ * Fine-tuning Phase: Further instruction-tuned on over 490k multi-turn conversational data to enable more instruction-following and context-aware responses.
94
+
95
+ ## Generic Capabilities on Vicuna Benchmark
96
+
97
+ The data is translated into traditional Chinese for evaluating the general capability.
98
+
99
+
100
+ <img src="./images/zhtw_vicuna_bench_chatgptbaseline.png" width="700">
101
+
102
+ 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.
103
+
104
+ | Language Model | Relative Score (%) |
105
+ |-------------------------------------|--------------------|
106
+ | GPT-4 | 102.59% |
107
+ | ChatGPT | 100.00% |
108
+ | **Taiwan-LLaMa v1.0** | 76.76% |
109
+ | Claude-Instant-1.2 | 74.04% |
110
+ | Llama2_Traditional_Chinese_13b_Chat | 56.21% |
111
+
112
+
113
+
114
+
115
+ ## How to deploy the model on my own machine?
116
+ 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.
117
+ ```bash
118
+ bash run_text_generation_inference.sh "yentinglin/Taiwan-LLaMa" NUM_GPUS DIR_TO_SAVE_MODEL PORT MAX_INPUT_LEN MODEL_MAX_LEN
119
+ ```
120
+
121
+ Prompt format follows vicuna-v1.1 template:
122
+ ```
123
+ 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:
124
+ ```
125
+
126
+ ## Setup development environment
127
+ ```bash
128
+ conda create -n taiwan-llama python=3.10 -y
129
+ conda activate taiwan-llama
130
+ pip install -r requirements.txt
131
+ ```
132
+
133
+
134
+ ## Citations
135
+ If you use our code, data, or models in your research, please cite this repository. You can use the following BibTeX entry:
136
+
137
+ ```bibtex
138
+ @inproceedings{lin-chen-2023-llm,
139
+ title = "{LLM}-Eval: Unified Multi-Dimensional Automatic Evaluation for Open-Domain Conversations with Large Language Models",
140
+ author = "Lin, Yen-Ting and Chen, Yun-Nung",
141
+ booktitle = "Proceedings of the 5th Workshop on NLP for Conversational AI (NLP4ConvAI 2023)",
142
+ month = jul,
143
+ year = "2023",
144
+ address = "Toronto, Canada",
145
+ publisher = "Association for Computational Linguistics",
146
+ url = "https://aclanthology.org/2023.nlp4convai-1.5",
147
+ pages = "47--58"
148
+ }
149
+
150
+ @misc{taiwanllama,
151
+ author={Lin, Yen-Ting and Chen, Yun-Nung},
152
+ title={Taiwanese-Aligned Language Models based on Meta-Llama2},
153
+ year={2023},
154
+ url={https://github.com/adamlin120/Taiwan-LLaMa},
155
+ note={Code and models available at https://github.com/adamlin120/Taiwan-LLaMa},
156
+ }
157
+ ```
158
+
159
+ ## Collaborate With Us
160
+ If you are interested in contributing to the development of Traditional Chinese 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.
161
+
162
+ ## License
163
+ The code in this project is licensed under the Apache 2.0 License - see the [LICENSE](LICENSE) file for details.
164
+
165
+ 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.
166
+
167
+ ## OpenAI Data Acknowledgment
168
+ 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.
169
+
170
+
171
+ ## Acknowledgements
172
+
173
+ 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.