I Distilled an LLM into two 287M encoders (GLiNER + multiple choice) for document extraction, can't match teacher. did i do something wrong?

A while ago I asked here how to turn ~5 million court decisions into structured graphs without running an expensive LLM on every document thanks for the advice . I went with the "small extractor + classifier" idea and it mostly works, but I'm stuck a bit below the LLM. And like I said last time, i'd