The Stanford team plagiarized a large model from Tsinghua University. The author apologized late at night. China’s large model can no longer be ignored.

Some time ago, the Stanford University Artificial Intelligence Institute (Stanford HAI) released a report stating that the United States is far ahead in the field of large models. The report pointed out that 61 well-known artificial intelligence models in 2023 came from US institutions, far exceeding the 21 in the EU and 15 in China.

Vinod Khosla, an early investor in OpenAI, also published an article on X last year saying that American open source models will be copied by China.

However, the large domestic model that has always been considered to be "catching up with the United States" has now become the target of plagiarism, and the plagiarized AI team is from Stanford University, which released the above report.

The Llama3-V open source model led by the Stanford AI team was found to be suspected of plagiarizing the domestic open source model "Little Steel Cannon" MiniCPM-Llama3-V 2.5 from Tsinghua University and Wall-Facing Intelligence, which instantly caused a stir in the AI ​​circle.

Under the real hammer, the Stanford team also had to apologize urgently.

As Li Dahai, CEO of Wall-Facing Intelligence, responded jokingly, this is a "method recognized by the international team." No matter how far we are from the top large models, domestic large models have reached a stage where they can no longer be ignored.

Let’s briefly summarize the timeline:

  • Stanford AI team releases Llama3-V, known as SOTA multi-modal large model
  • Netizens questioned that the model copied the domestic wall-facing smart MiniCPM-Llama3-V2.5
  • Questioning evidence emerged, Llama3-V author staged "deleting the database and running away"
  • Face Wall Intelligence official plagiarism, issued a statement late at night
  • Llama3-V author formally apologizes, netizens have different opinions

Plagiarizing the wall-facing intelligent "Small Steel Cannon", the Stanford AI team staged "Delete the database and run away"

Recently, a Stanford AI team announced that it only costs $500 to train a SOTA multi-modal large model that surpasses GPT-4V.

But soon, an X user @yangzhizheng1 pointed out that the model structure and code used in this project are surprisingly similar to the MiniCPM-Llama3-V2.5 released by Wallface Intelligence not long ago.

To this end, X user @yangzhizheng1 also released corresponding questioning evidence.

Evidence one:

The model structure and code of Llama3-V and MiniCPM-Llama3-V 2.5 are almost copy-paste level similar. The difference is probably that they have changed the vest – the variable names have been changed.

It's like the same dress, but with buttons of different colors. Do you think it's a coincidence?

Evidence two:

When the authors of Llama3-V were asked why they could use the MinicPM-Llama3-V2.5 tokenizer that had not yet been released in advance, they explained that they were using the previous generation MinicPM-V-2 project of Wall-Facing Intelligence.

However, some media sought confirmation from Wallface Intelligence officials. In HuggingFace, the MiniCPM-V2 and MiniCPM-Llama3-V 2.5 word segmenters are two files respectively, and the file sizes are completely different.

What's more, the tokenizer of MiniCPM-Llama3-V 2.5 is composed of the Llama3 tokenizer plus the special token of the MiniCPM-V series model.

Considering that MiniCPM-V2 was released earlier than Llama3, it is theoretically impossible for it to include the Llama3 tokenizer technology that has not yet been disclosed.

Evidence three:

What's even more outrageous is that when the author of the llama3-V project faced users' doubts and saw that something was not going well, he simply staged a good show of "deleting the library and running away."

Even the project page on GitHub has been removed, which can be called deceptive version 2.0.

The Hugging Face address is as follows. Currently, when we open the page, we can only see "404".

https://huggingface.co/mustafaaljadery/llama3v/commit/3bee89259ecac051d5c3e58ab619e3fafef20ea6

This is not over yet, more evidence is emerging:

X User @yangzhizheng1 said that if Gaussian noise (parameterized by a single scalar) is added to the checkpoint of MiniCPM-Llama3-V 2.5, the resulting model will be carved out of the same mold as Llama3-V.

Not only that, this model can also recognize the profound ancient writings of the Warring States Period such as "Tsinghua Slips", and the errors are exactly the same. In the official words of Wall-Facing Intelligence:

Not only are they right, they are also wrong.

You must know that this ancient writing data was obtained by scanning and manually annotating the Tsinghua bamboo slips collected by Tsinghua University over several months. It has never been made public.

So how did the Stanford AI team get it out of thin air?

It can be said that Wallface Intelligence’s late-night statement on June 2 can be regarded as a complete plagiarism from the Stanford AI research team.

Until early this morning, Siddharth Sharma and Aksh Garg, two authors of the Stanford Llama3-V team, formally apologized to the wall-facing MiniCPM team for this academic misconduct on social platform X, saying that all Llama3-V models would be removed.

Do top students from famous schools also plagiarize? China’s open source big models are catching up

An important reason why this matter caused a stir on the Internet is that the background of the plagiarist is so glamorous.

Public information shows that Siddharth Sharma and Aksh Garg are both undergraduate students in the Department of Computer Science at Stanford University and have published many papers in the field of machine learning. Among them, Siddharth Sharma has interned at Amazon for a period of time and is currently mainly engaged in AI and data-related work.

Aksh Garg's internship resume is rich, covering well-known organizations such as SpaceX, Stanford University and California Institute of Technology.

As for Mustafa Aljadery, who is called the "code porter" by the two above-mentioned authors, he is a graduate of the University of Southern California. After public opinion fermented, the X account has been set to private status.

Sharp-eyed netizens did not accept the apology statement from the Stanford Llama3-V team.

For example, the

Christopher David Manning, director of the Stanford AI Laboratory, also stood up to condemn this plagiarism and praised MiniCPM, an excellent Chinese open source model.

However, there are also netizens who hold the attitude of "forgive others when you have to give them mercy" and encouraged them leisurely:

Openness and honesty are very important values ​​in the technology world, and I look forward to your new work.

Google DeepMind researcher Lucas Beyer said that China’s open source large models have good models like MiniCPM, but the international community has not given them enough attention…

The Wall-Facing Intelligence team also responded to this matter yesterday.

Li Dahai, CEO of FaceWall Intelligence, said: "Technological innovation is not easy. Every job is the result of the team's day-and-night efforts and a sincere contribution to technological progress and innovative development around the world with limited computing power.

We hope that the good work of the team will be noticed and recognized by more people, but not in this way. "

Liu Zhiyuan, chief scientist of Wall-Facing Intelligence, also posted on Zhihu, saying that this incident proved the international influence of China’s innovative achievements from another perspective, emphasizing the importance of open source sharing and respect for the spirit of originality.

It has to be said that this plagiarism drama in the AI ​​​​circle is a textbook explanation of "innovation is not easy, it must be done and cherished, academic integrity is everyone's responsibility."

You know, if you imitate the shape of the code, you can't copy the original grace.

In fact, since last year, China's large models have been open sourced like mushrooms after a spring rain. They have transformed from beneficiaries to contributors and are not stingy in providing more outstanding open source results to the world.

From giants such as Alibaba and Tencent to wall-facing intelligence, AI start-ups such as Zhipu AI and Kunlun Tiangong are also active members of the open source community, contributing to the development of China's large-scale models.

We also hope that this spring breeze of openness and sharing will blow stronger.

Just as Li Dahai, CEO of Face Wall Intelligence, calls for everyone to work together to build an open, cooperative and trusting community environment. Only by working together can the world become a better place with the arrival of AGI!

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo