Two authors behind a Stanford University AI project have apologized to the Chinese team behind open-source AI model MiniCPM-Llama3-V 2.5 after social media users in China outed the former for having plagiarized the latter model, which was developed by Tsinghua University and ModelBest Inc. The incident sparked widespread discussion on the Chinese internet.
Two authors of the Stanford Llama3-V team, Siddharth Sharma and Aksh Garg, today apologized to the MiniCPM team on social media site X for their academic misconduct, announcing they would withdraw the Llama3-V model from use.
Why it matters: Members of the Stanford team acknowledged they had plagiarized Tsinghua University and ModelBest Inc’s work.
Details: “We sincerely apologize to the authors of MiniCPM for our failure to verify the originality of Llama3-V,” Aksh and Siddharth wrote on X today. “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.”
- On May 29, a Stanford AI team claimed online that they could train a multi-modal large model surpassing GPT-4V for only $500, according to local media outlet Quantum Bit. Subsequently, social media users discovered that the team’s Llama3-V model used a model structure and code highly similar to ModelBest’s MiniCPM-Llama3-V2.5, with partial variable name modifications. Llama3-V also features the same tokenizer as MiniCPM-Llama3-V 2.5, including the latter’s newly defined special symbols.
- On June 2, Chinese firm ModelBest Inc confirmed that the Stanford large model project Llama3-V, similar to MiniCPM, is able to identify Qinghua Jian ancient Chinese characters from China’s Warring States Period. Remarkably, the matches also shared identical mistakes. This character data, obtained by the research team through months of scanning and manual annotation of each Qinghua Jian character, had not been publicly disclosed, confirming the act of plagiarism.
- “While it’s good to be recognized by international teams, we believe in building a community that’s open, cooperative, and trustworthy,” ModelBest Inc CEO Li Dahai stated. “We want our team’s work to be noticed and respected, but not in this manner.”
- ModelBest Inc’s chief scientist, Liu Zhiyuan, who is also a tenured associate professor at Tsinghua University, told an Yicai reporter that the rapid development of artificial intelligence relies on the global sharing of algorithms, data, and models. “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,” he said.
- Christopher David Manning, the director of the Stanford AI Laboratory, also issued a statement condemning the act of plagiarism and praising the Chinese open-source model MiniCPM.
Context: Founded in August 2022, ModelBest Inc secured a new round of financing worth hundreds of millions of RMB in April. Huawei’s Hubble Technology Venture Capital led the investment, with participation from Chunhua Capital, Beijing Artificial Intelligence Industry Investment Fund, and Chinese Quora-like platform Zhihu. In February, ModelBest Inc launched the open-source model MiniCPM.
Related
