An AI team from Stanford University recently issued an apology for alleged plagiarism concerning their model, Llama3-V. This team, comprised of three Stanford students, released the open-source model, which soon faced scrutiny for its striking similarities to an existing model by the Chinese company, Beijing MindWall Intelligence. This model, named MiniCPM-Llama3-V 2.5, had already gained recognition for its innovative features.
On June 3, Llama3-V team members Siddharth Sharma and Aksh Garg publicly apologized on social media, announcing their decision to withdraw the model from circulation. They clarified that although they promoted the model, the original code's authorship belonged to Mustafa Aljadery. After reaching out to Aljadery to clarify the allegations, the team lost communication with him and opted to issue an apology.
Aljadery's social media account is currently restricted to "friends only," and the Llama3-V model has been removed from platforms like Hugging Face, although some descriptions remain on its Medium page alongside the apology announcement.
Founded in August 2022, Beijing MindWall Intelligence has developed significant products, including the ModelForce platform and the CPM model. The company secured multi-million funding in April, led by Huawei, with additional investments from other firms, including Zhihu, which remains a strategic partner.
Before the Llama3-V team's apology, MindWall's co-founder and CEO, Li Dahai, highlighted evidence via social media indicating that Llama3-V could recognize ancient Chinese characters, similar to MiniCPM's capabilities. He pointed out that MiniCPM's dataset had not been publicly released, reinforcing claims of intellectual infringement. Li expressed dissatisfaction, stating, "We wish for our team's accomplishments to gain recognition, but not in this manner... We call for the creation of a cooperative and trust-building community."
The controversy began on May 29 when the Stanford team announced their model, claiming it could be trained for just $500 to achieve state-of-the-art multi-modal efficacy, rivaling established models like GPT-4V and Gemini Ultra. The authors, with backgrounds at notable companies such as Tesla and Amazon, saw their announcement garner over 300,000 views and a rapid rise in Hugging Face's model rankings.
Despite the initial enthusiasm, skepticism grew within the open-source community. Critics accused the Llama3-V team of rebranding MindWall's MiniCPM without proper attribution. The team defended themselves, asserting they had utilized MiniCPM's tokenizer, having started their project before MiniCPM's public release. Nonetheless, users in the community highlighted significant code and model structure overlaps, with evidence provided by a Reddit user.
As the controversy escalated, suggestions to alert Stanford University about the issue circulated on social media, with some commenters expressing concern over potential damage to the open-source community caused by the team's actions without adequate credit.
In a separate statement, Christopher David Manning, director of Stanford's AI lab, condemned the plagiarism, acknowledging MiniCPM as a valuable open-source contribution. Liu Zhiyuan, Chief Scientist at MindWall and a tenured professor at Tsinghua University, also reflected on the incident, emphasizing the need for confidence and proactive engagement as research evolves in the age of artificial general intelligence (AGI).
Ranking
