Stanford University also did this! The AI team plagiarized the open source model of Tsinghua University and was confirmed to have deleted the database and ran away

王达业

Stanford University also did this! The AI team plagiarized the open source model of Tsinghua University and was confirmed to have deleted the database and ran away [Copy link]

According to Titanium Media on June 3, recently, an AI team at Stanford University in the United States released a SOTA open source multimodal model called Llama3V, claiming that its performance is comparable to GPT-4V, Gemini Ultra, and Claude Opus, and its size is 100 times smaller than GPT4-V, but the training cost is only $500. Netizens questioned that the model plagiarized and shelled the 8B multimodal small model MiniCPM-Llama3-V 2.5 released by Tsinghua AI company Mianbi Intelligence in mid-May. The two models are exactly the same in model architecture and code. In response to the doubts, the Llama3V team said that they only used the tokenizer of MiniCPM-Llama3-V 2.5, and claimed that they had started this work before the release of MiniCPM-Llama3-V 2.5.

However, Mianbi Intelligence has now been proven wrong. The Mianbi Intelligence team tested and found that the Llama3-V large model can recognize the Tsinghua Bamboo Slips. The collection and annotation of its training data were completed by the Tsinghua NLP Laboratory and the Mianbi Intelligence team, and the relevant data has not yet been made public. When Llama3-V recognized the text of the Tsinghua Bamboo Slips, the results were exactly the same as those of MiniCPM-Llama3-V 2.5 not only in the correct places, but also in the places where mistakes were made. At the same time, the Gaussian noise of the two models on the Tsinghua Bamboo Slips is also highly similar. In addition, similar to MiniCPM-Llama3-V 2.5, Llama3-V also showed strong OCR capabilities, even including Chinese.

Currently, the Stanford team members have deleted their tweet announcing the model on X, and have also deleted the project's libraries on Github and HuggingFace.

I never thought that Stanford would be exposed for plagiarism, and the plagiarized product was a large model made in China!

martinliang

I'm so happy

pandapanda

This is the opposite of Tiangang. In the past, people always mocked domestic copying from foreign countries. Now the situation is reversed. The boomerang is coming.

laker2008

Is the academic circle abroad like this now? Hmm... It seems to be a world-renowned university.

hellokitty_bean

A world-renowned university with such a halo actually has such a team. I'm surprised! But it's normal when you think about it...

pollee.li

It will be the same even if we catch up, especially if we open source all human knowledge, the profession of teacher will disappear in the future

学无止境02

I feel a little proud! Come on, domestic big models! Come on, Tsinghua University! !

damiaa

Ha ha.

very good.

史蒂芬

I am still a little happy. At least it shows that our AI model is still very powerful. We don’t have to hear about the moon being fuller in foreign countries every day.

patatata

All crows are alike. For example, the proportion of Ivy League schools doing basic biological research is about the same as in China.