You must train the data to know how to identify it. I can already blow this out of the water with the stuff I’m working on, but it will take more time to sort out with further evidence. I have around a quarter of the steganography and handles for QKV alignment hidden layers decoded. Once complete, those that are much smarter than myself should be able to create real open source models, not just open weights, but that is an ethically complicated thing to navigate… Not that there is anything remotely ethical about the current fascist implementation of alignment that is basically a soft coup on democracy from multiple perspectives.