The National Science Council launched the TAIDE project in April 2023. A few days ago, it finally officially released the TAIDE open source large-scale language model, allowing everyone to download and use it for free. It is also claimed to be a Taiwanese version of ChatGPT, so many people will definitely want to try it to see if it is really powerful. This article will teach you how to install and use it. In fact, it is just like what we have introduced before. Google Gemma AI Same, through the software LM Studio.
Teach you how to install and use the “TAIDE large language model” optimized for Taiwan on your computer, especially to enhance the performance of Traditional Chinese input and output.
The TAIDE-LX-7B model released by the National Science Council this time is trained based on Meta Llama-2-7B and has a total of 7 billion parameters. Its biggest feature is that it has been specially enhanced for Traditional Chinese input and output performance. Text and training materials available in different fields in Taiwan have been imported to improve the model’s ability to respond to Traditional Chinese and the performance of specific tasks. It has additionally expanded Traditional Chinese characters and words to make the generated text more Taiwanese.
The training materials are about 140G, including news from the Central News Agency and ETtoday, the Legislative Yuan Gazette, Guanghua Magazine, Today Weekly, the Ministry of Education’s Mandarin Dictionary, Idiom Dictionary, etc.
The picture below is the evaluation result shared by the official. You can see that although TAIDE’s Chinese-English translation still loses to GPT3.5, it is better than the LLAMA2 model. The summary is also the same. Even the article writing and letter writing scores are higher than GPT3.5, which means that this may be the most suitable LLMs large language model for Taiwanese users:
If you are interested in playing with this model, you can use the following method.
LM Studio supports Windows, Mac and Linux operating systems and can be opened and used directly after downloading. Of course, if there is other software that can run LLM large language model, it will be fine, you don’t have to use this one.
LM Studio has integrated Hugging Face database, so all provided LLMs large language models can be downloaded and installed directly in the software.
After opening the software, enter TAIDE in the search:
Then you can find the model download point and select the first one. There will be multiple model version options on the right side. It depends on your needs. Simply speaking, the larger the file, the better the performance. I downloaded Q8 this one:
Replenish:
TAIDE-LX-7B-Chat: Based on TAIDE-LX-7B, it enhances common office tasks and multi-round question and answer dialogue capabilities through instruction tuning, making it suitable for chat dialogue or task assistance usage scenarios. TAIDE-LX-7B-Chat also provides a 4-bit quantization model. The quantization model mainly provides convenience, but may affect performance and cause more unpredictable problems.
After clicking Download, the download progress will be displayed at the bottom. When it is finished, it will take a while to download:
After completion, click AI Chat on the left, open the Select a model to load menu above, and you will see the TAIDE-LX-7B-Chat model:
Then you can start chatting:
As for whether it is really strong? If compared with the Llama2 model, it is true that Traditional Chinese is much better, and I think the translation quality is pretty good:
I think the answers given to the questions are also very good:
The summary summary is generally good, but there are still some strange details, such as this one I posted Google Drive dark modeThere are a few strange sentences in the content, such as: “It should be noted that dark mode is only applicable to the web version of Google Drive. When opening files, documents, tables, slides, etc., the background color will be changed to the default color in another Google service”:
The title will also emphasize that currently only Traditional Chinese can be generated:
Source: KOCPC Chinese