Definitely understand the current model's shortcomings that people have mentioned.

I actually think this is largely in part due to how the current model + training process is designed.

I have some ideas for improving things on this front, will give it a try and push an update soon! :)