Stanford Team Apologises to Chinese AI Model Developers for Plagiarism
Updated: Jun 7
Two Stanford University AI researchers have apologised for plagiarising the Chinese open-source AI model MiniCPM-Llama3-V2.5. The Stanford team said that they failed to validate the authenticity of its Llama3-V model. ModelBest Inc acknowledged plagiarism by pointing up similarities between Llama3-V and MiniCPM, such as the usage of the same tokenizer and same errors in detecting ancient Chinese characters.
The apology came after social media users in China accused the Stanford researchers of plagiarising the MiniCPM model, which was initially developed by Tsinghua University and ModelBest Inc. This incident provoked extensive discussion on the Chinese internet.
Siddharth Sharma and Aksh Garg, members of the Stanford Llama3-V team, turned to social media site X to apologise for their academic malfeasance. They announced that the Llama3-V model would be discontinued.
This apology is significant since members of the Stanford team admitted to plagiarising Tsinghua University and ModelBest Inc's work.
In their statement on X, Aksh and Siddharth wrote, "We sincerely apologise to the authors of MiniCPM for our failure to verify the originality of Llama3-V. Mustafa, who wrote the code, described exciting extensions that we promoted without knowing about the prior work by OpenBMB (founded by Tsinghua University and ModelBest Inc). We take full responsibility for this oversight. We've removed all references to Llama3-V in respect to the original authors."
The Stanford researchers sparked controversy when they stated online on May 29 that they could train a multi-modal big model that outperformed GPT-4V for only $500, according to local media outlet Quantum Bit. However, social media users quickly noticed that the team's Llama3-V model was strikingly similar to ModelBest's MiniCPM-Llama3-V2.5, with just minor variable name changes. Furthermore, Llama3-V also utilised the same tokenizer as MiniCPM-Llama3-V2.5, including the newly defined special symbols.
On June 2, ModelBest Inc, a Chinese firm, confirmed that the Stanford large model project Llama3-V, similar to MiniCPM, was capable of identifying Qinghua Jian ancient Chinese characters from China's Warring States Period. Notably, the matches between the two models also shared identical mistakes. This character data, obtained through months of scanning and manual annotation, had not been publicly disclosed, further confirming the act of plagiarism.
ModelBest Inc's CEO, Li Dahai, expressed his disappointment, stating, "While it's good to be recognised by international teams, we believe in building a community that's open, cooperative, and trustworthy. We want our team's work to be noticed and respected, but not in this manner."
Liu Zhiyuan, the chief scientist of ModelBest Inc and a tenured associate professor at Tsinghua University, emphasised the importance of global sharing in the rapid development of artificial intelligence. He commented, "Two out of the three members of this Llama3-V team are merely undergraduate students at Stanford University, and they have a long journey ahead. If they can acknowledge their mistakes and make amends, it would be a great virtue."
Christopher David Manning, head of the Stanford AI Laboratory, criticised plagiarism and lauded the Chinese open-source model MiniCPM.
In this backdrop, ModelBest Inc, which was created in August 2022, just closed a new round of financing totaling hundreds of millions of RMB. Huawei's Hubble Technology Venture Capital spearheaded the investment, which also included Chunhua Capital, the Beijing Artificial Intelligence Industry Investment Fund, and the Chinese platform Zhihu. ModelBest Inc launched MiniCPM, an open-source model, in February.
Two authors from a Stanford University AI project have apologised for plagiarising the Chinese open-source AI model MiniCPM-Llama3-V2.5.
The Stanford team acknowledged their failure to verify the originality of their Llama3-V model.
ModelBest Inc confirmed the act of plagiarism by highlighting the similarities between Llama3-V and MiniCPM, including the use of the same tokenizer and identical mistakes in identifying ancient Chinese characters.
Source: TECHNODE