Did Stanford’s Llama3‑V Steal MiniCPM‑Llama3‑V 2.5? A Deep Dive into the Evidence

Background

On 29 May 2024 a Stanford AI team announced Llama3‑V , a multimodal large‑language model claimed to be trainable for $500 and to surpass GPT‑4V, Gemini Ultra and Claude Opus. Shortly after, a community member reported that the model’s architecture, source files and configuration are almost identical to the open‑source project MiniCPM‑Llama3‑V 2.5 from the Chinese startup Mianbi Intelligent (面壁智能). The MiniCPM repository is available at https://github.com/OpenBMB/MiniCPM-V and the Llama3‑V repository at https://github.com/mustafaaljadery/Llama3-V.

Technical Evidence

1. Architecture and Code

Side‑by‑side screenshots show that the model definition files in Llama3‑V match MiniCPM‑Llama3‑V 2.5 line‑by‑line, differing only in variable names. Configuration files also share the same structure, with only minor changes to image‑slice, tokenizer, resampler and data‑loader settings.

2. Tokenizer Reuse

Llama3‑V claims to use a tokenizer from an earlier MiniCPM‑V project. However, the tokenizer files in MiniCPM‑Llama3‑V 2.5 differ in size and content from those released with MiniCPM‑V 2, and the Llama3‑V tokenizer appears to be derived from the MiniCPM‑Llama3‑V 2.5 version that was not publicly available at the time of Llama3‑V’s release.

3. Repository Removal

After the plagiarism accusations, the original issue thread on the MiniCPM repository (issue #196) was deleted, and the Llama3‑V HuggingFace page returned 404 errors, indicating that the project files were taken down.

Reproduction Experiments

Running Llama3‑V with the MiniCPM‑Llama3‑V 2.5 codebase and configuration produced successful inference without modification.

Renaming the variable names in the MiniCPM checkpoint to match Llama3‑V allowed the model to load and run under the Llama3‑V code.

Adding a small amount of Gaussian noise (single‑scalar parameter) to the MiniCPM‑Llama3‑V 2.5 checkpoint yielded a model whose behavior was nearly indistinguishable from Llama3‑V.

On a test set of 1,000 bamboo‑slip images, the two models produced identical predictions for 87 % of the samples, far higher than any other model pair evaluated.

The error analysis showed that Llama3‑V and MiniCPM‑Llama3‑V 2.5 made 236 and 194 incorrect predictions respectively, with 182 errors overlapping, indicating a highly similar error distribution.

Both models exhibited the same failure patterns in the proprietary “WebAgent” mode, a feature internal to MiniCPM‑Llama3‑V 2.5.

Responses

The Stanford authors initially argued that their work pre‑dated MiniCPM and that only the tokenizer was reused. In a later Medium statement they acknowledged that their architecture closely resembled MiniCPM‑Llama3‑V 2.5 and that they had removed the original model files.

"We realize our architecture is very similar to OpenBMB’s MiniCPM‑Llama3‑V 2.5, which implemented the idea before us. We have deleted the original model."

MiniCPM‑Llama3‑V 2.5’s lead developer, Li Dahai, issued a formal request for an explanation from the Stanford team.

References

https://github.com/OpenBMB/MiniCPM-V/issues/196

https://github.com/mustafaaljadery/Llama3-V

https://www.reddit.com/r/LocalLLaMA/comments/1d6f1f3/Llama3-V_project_is_stealing_a_lot_of_academic/

https://aksh-garg.medium.com/llama-3v-building-an-open-source-gpt-4v-competitor-in-under-500-7dd8f1f6c9ee

Code example

往
期
推
荐
1、
Spring Boot + URule 实现可视化规则引擎，太优雅了！
2、
Git严重漏洞，远程执行代码，Mac和Windows通杀！
3、
互联网人为什么学不会摆烂
4、
我认为这几款VSCode插件很好用，助力我半吊子前端进阶之路
5、
相比高人气的Rust、Go，为何 Java、C 在工具层面进展缓慢？
6、
运营28年的软件即将关停：鼎盛期用户数超1亿、腾讯都得叫声前辈
点
分
享
点
收
藏
点
点
赞
点在看