Sorry, there is a fundamental difference with modification when you get into software, server-based logic, and machine learning.
It's a bit like when people used to modding their own hardware tools encounter a GPU or a RAM card that won't fit into a PCI slot, so they "cut it to fit", and then are surprised it doesn't work.
When you modify a car physically with spikes or to make more engine noise, or to have looser suspension you are dealing with isolated newtonian environments.
When you tweak software that involves massive machine learned models, you can cause absolutely unpredictable horrific consequences. There are good reasons to restrict this.
It's a bit like when people used to modding their own hardware tools encounter a GPU or a RAM card that won't fit into a PCI slot, so they "cut it to fit", and then are surprised it doesn't work.
When you modify a car physically with spikes or to make more engine noise, or to have looser suspension you are dealing with isolated newtonian environments.
When you tweak software that involves massive machine learned models, you can cause absolutely unpredictable horrific consequences. There are good reasons to restrict this.