I just don’t see how using anything out of this model could be useful for fine-tuning. It will barely follow instructions or guidance, completely ignores reference code and documentation that could get output in line; the system message you are trying to minimize that can’t be placed at all.
“Here’s your revised snippet from your 2000-line project, with all the exception hierarchy I couldn’t understand about deleting variables stripped, and all the function calls replaced with my pre-training. You like ‘gpt-3.5’ and everything made non-working right? Oh, and for the libraries you already made extensive use of, here’s how you can pip install”…
I think there’s just simply too much context junk inserted after your input for the gpt-4o base inside to still pay attention, so it reverts to what its small model (and the fine tune model) is powered by instead of emergent intelligence that scales: post-training.