Fine-tune evaluation

Measuring what a fine-tune actually changed, usually by scoring base and tuned models on the same task set with an external grader such as a compiler or schema validator.