Fine-tuned models, while effective for common scenarios, are brittle and break with less frequent use cases or when input format varies from training data.
Imagine a fine-tuned model complemented by another model when it requires knowledge out of its scope.The Leeroo Math 7B model exemplifies this architecture. When it doesn't know how to solve a problem, it calls GPT4.
https://huggingface.co/leeroo/LeerooDedicated-Math-7b