[License-discuss] Pre-Advice for ModelGo Licenses Review

[Moming Duan]

Thanks for the insights from David and Josh. I agree that we must ensure the responsible use of open-source works. However, I see a key dilemma in model sharing: there is currently no OSD-compatible model license available for users. Historically, training LLMs has been extremely costly, making it accessible only to big tech companies with legal teams that impose restrictions through proprietary licenses. However, as hardware and software evolve, more LLMs and their derivatives are emerging—currently, there are 1.4 million models on Hugging Face. Some of these models are joining the open-source community to advance AI technology. 

At present, open-source models primarily adopt licenses like Apache-2.0 and MIT, which were drafted before the GPT era and have limitations in model publishing. For example, licensees can bypass attribution requirements through knowledge transfer—training a new model based on the output of an existing one, as seen with DeepSeek training on ChatGPT-generated content. This is why I advocate for the ModelGo License: a set of reusable, model-specific licenses.

In my view, openness is crucial, and the contributions of open-source model projects should be respected. It is unfair that proprietary LLMs, like Llama 2, explicitly prohibit their use for improving other models, while open-source models can be distilled to enhance proprietary LLMs—without even requiring attribution.I have submitted the ModelGo License for OSI review, and I hope it can contribute to promoting a more open AI ecosystem.


Best,
Moming