Stanford AI Project Authors Apologize for Plagiarizing Chinese Large Model and Pledge to Remove Related Models
Recently, the Llama3-V open-source model led by the AI team at Stanford University was confirmed to have plagiarized the “Mini Cannon” MiniCPM-Llama3-V 2.5 open-source model released by Tsinghua University and Mianbi Intelligence in China. This has sparked heated discussions online.
In the latest development, two authors of the Stanford Llama3-V team, Siddharth Sharma and Aksh Garg, formally apologized to the Mianbi MiniCPM team on social media for this academic misconduct and indicated they would remove all Llama3-V models, per Chinese media Yicai.
Aksh Garg stated, “First of all, we want to apologize to the original authors of MiniCPM. Siddharth Sharma, Mustafa, and I released Llama3-V together, with Mustafa writing the code for this project, but he has been unreachable since yesterday. Siddharth Sharma and I were mainly responsible for helping Mustafa promote the model. We both reviewed the latest papers to verify the novelty of this work, but we were not informed or aware of any previous work by OpenBMB (a large-scale pre-training language model library and related tools initiated by the Tsinghua team). We apologize to the authors and are disappointed that we did not strive to verify the originality of this work. We take full responsibility for what happened and have removed Llama3-V, again, we apologize.”
Additionally, Christopher David Manning, the director of the Stanford Artificial Intelligence Laboratory, also posted a condemnation of this plagiarism and praised the MiniCPM, a Chinese open-source model.
The incident originated on May 29th, when a Stanford AI team advertised online that a SOTA multimodal large model surpassing GPT-4V could be trained for just $500. Subsequently, netizens discovered that the model structure and code used by the Llama3-V model were extremely similar to the MiniCPM-Llama3-V2.5 released by Mianbi Intelligence not long ago, with only some variable names modified. Llama3-V also has the same tokenizer as MiniCPM-Llama3-V 2.5, including the latter’s newly defined special symbols.
Late on the night of June 2nd, the Mianbi Intelligence team confirmed that the Stanford large model project Llama3-V, like MiniCPM, could recognize the “Tsinghua Simple” ancient Warring States script, “not only getting it exactly right, but also making the exact same mistakes.” This ancient script data was manually annotated by the research team after months of scanning word by word from the Tsinghua Simple, and was not publicly available, confirming the fact of plagiarism.
Mianbi Intelligence CEO Li Dahai stated, “We deeply regret this incident. On one hand, we feel that this is also a way of being recognized by international teams, on the other hand, we call for everyone to build an open, cooperative, and trustworthy community environment.” “We hope that the good work of the team will be noticed and recognized by more people, but not in this way.”
Mianbi Intelligence’s Chief Scientist and Associate Professor at Tsinghua University, Liu Zhiyuan, stated that the rapid development of artificial intelligence cannot be separated from the open-source sharing of global algorithms, data, and models, allowing people to continually advance on the shoulders of SOTA. Mianbi’s open-source MiniCPM-Llama3-V 2.5 uses the latest Llama3 as the language model base. The cornerstone of open-source sharing is adherence to open-source protocols, trust in other contributors, and respect and tribute to the achievements of predecessors, which the Llama3-V team undoubtedly severely damaged. They have deleted the library on Huggingface after being questioned, and two of the three team members are just undergraduates at Stanford University, with a long way to go in the future. If they can correct their mistakes, it will be a great virtue.
Beijing Mianbi Intelligence Technology Co., Ltd. was established in August 2022. In April of this year, Mianbi Intelligence completed a new round of multi-billion yuan financing, led by Huawei Hubble, with Chunhua Venture Capital, Beijing Artificial Intelligence Industry Investment Fund, and others following. Zhihu continued to follow and support as a strategic shareholder. In February of this year, after Mianbi Intelligence released the open-source model MiniCPM, it launched the MiniCPM 2 series of edge models. Li Dahai stated that promoting the landing of large models on the edge is one of the key tasks of Mianbi at present.
