Fengse is sent from the concave non-si qubit
| public account QbitAI
Finally, Tang Jie's team from Tsinghua University also made a move.
On the same day that GPT4 was released, Professor Tang announced on Weibo:
ChatGLM, a dialogue robot based on a large model with hundreds of billions of parameters , is now open for internal testing by invitation.
According to reports, ChatGLM is specially optimized for Chinese , which can be said to be friendly to Chinese people~
It seems that writing blog outlines, creating scripts, solving math problems, writing codes, and even playing the role of Gao Leng Yujie , it is very good at it:
△ Write a blog outline introducing ChatGLM
△ Cosplay chat as Gao Leng Yujie
Qubit was fortunate enough to get the internal test quota, and will conduct a wave of actual tests later.
Meanwhile, another announcement was made:
In order to better promote the development of large-scale model technology together with the community, Tang Jie's team also open sourced the bilingual ChatGLM-6B model containing 6.2 billion parameters.
Its biggest feature is that it can be deployed on an ordinary computer , as long as there is a 2080Ti.
Lets come look.
Can write speeches in classical Chinese, know the latest news of He Yuming, but...
Let's look at ChatGLM first, it is a new member of Tang Jie's large model family that was born after half a year.
The alpha version is called QAGLM (full name qa-glm-v0.7).
Try out the effect.
First of all, let's let it talk about the difference between it and ChatGPT.
It points out its own research and development institutions, and the characteristics of being good at Chinese.
Then, let it write a thank you speech in classical Chinese.
How to say, although there is an extra "yu" and an inexplicable traditional character, it is still readable, and the parallel sentences used have enhanced the momentum.
Then, we threw the beginning of the article about Thunder in Silicon Valley a few days ago to it, and let it have a title.
It feels pretty good, at least a few key pieces of information have been captured.
Unfortunately the thesis challenge didn't pass , and when we threw the link to the GLM-130B at it to briefly recap the topic, it wasn't talking about this article at all.
It is almost as good as the operation of ChatGPT Hu Zou's reference (manual dog head).
Next, test its mathematical ability.
This elementary school word problem is no problem:
But the chicken and the rabbit are in the same cage, it is difficult for it, and in the end it even calculates a negative number ==
In terms of programming, you can also solve simple algorithm problems.
What about the ability to summarize information? We have given a requirement in English, it is not difficult:
The result is correct:
It should be noted that currently ChatGLM can only conduct a maximum of 5 rounds of dialogue in each round, and each time a maximum of 1,000 words can be entered.
It has a good grasp of new information. It knows that the current CEO of Twitter is Musk, and it also knows that He Yuming returned to academia on March 10, but it has not yet discovered that GPT-4 has been released .
And, the current response speed is still very fast. No matter what the question is, the answer is correct or not, and the answer can be given within a few seconds.
Finally, Qubit also let it have a cosplay to see how good it is to coax a girlfriend:
Emmm, although I was a bit upright, "I" really lost my anger after hearing this passage.
So, the above are our test results, what do you think?
Based on a base model with 130 billion parameters
According to the official introduction, ChatGLM refers to the design ideas of ChatGPT, injects code pre-training into the 100 billion base model GLM-130B , and realizes human intention alignment through supervised fine-tuning and other technologies (that is, to make the machine's answer conform to human values, human expect).
The background of this GLM-130B is worth talking about.
It is a large-scale Chinese-English pre-training language model jointly developed by the Knowledge Engineering Laboratory (KEG) of Tsinghua University and Zhipu AI. It has 130 billion parameters and was officially released in August last year.
Unlike the architectures of BERT, GPT-3, and T5, GLM-130B is an autoregressive pre-training model that includes multiple objective functions.
Its advantages include:
Among the 30 world mainstream large-scale model evaluations reported by Stanford, GLM-130B has also become the only selected model in Asia .
And got good results:
For example, it is close to or equal to GPT-3 175B (davinci) in terms of accuracy and maliciousness indicators, and the robustness and calibration error are in all pedestal models with a scale of 100 billion (as a fair comparison, only fine-tuning models without instruction prompts are compared) It is also remarkable.
And at a recent meeting of CCF, an audience asked: Why was ChatGPT not born in China? Is it that we didn't pay attention to this matter?
The guests moved out the GLM-130B (it was also selected for ICLR'23).
Now, GLM-130B has finally been put into "big use".
Regarding the internal test, Tang Jie's team said that the scope will be gradually expanded in the future, and friends who are interested can wait for a while.
The reduced version of 6 billion parameters is also open source
In addition to this chat robot ChatGLM, Tang Jie's team also open sourced the "shrunk version" ChatGLM-6B of GLM-130B this time.
△ GitHub has won nearly 2k stars
ChatGLM-6B uses the same technology as ChatGLM, and has the functions of Chinese question and answer and dialogue.
The features are as follows:
Of course, the disadvantage is that the capacity is only 6 billion, its model memory and language ability are weak, it is not good at logic problems (such as mathematics, programming), and multiple rounds of dialogue may cause context loss and misunderstanding.
But its main feature is a low threshold, which can be used for reasoning on a single 2080Ti, and the hardware requirements are not high.
Therefore, as long as you are interested, you can download it for a try, both for research and (non-commercial) application development.
Portal:
https://chatglm.cn/
https://github.com/THUDM/ChatGLM-6B
Reference link:
[1]https://weibo.com/2126427211/MxlsQ6w4A#repost
[2]https://chatglm.cn/blog?continueFlag=d70d7590143c950d12ac7283214d879d
—End— _ _