I think you could say something remarkably similar about just releasing bytecode as well and I think most people would call foul at that. I don't think it's so

I think you could say something remarkably similar about just releasing bytecode as well and I think most people would call foul at that. I don't think it's so cut and dry. This isn't entirely about being a beginner or not either. Full fine-tuning without forgetting does really want the training data (or something that is a good replacement). You can do things like LoRa but, depending on your use case, it might not work.