Did the Model Improve, or Just the Score? A Developer’s Guide to LLM Evaluation

A model can extract the right information and still score zero because its answer differs from a benchmark’s reference. Fine-tuning can raise that score, but does the improvement carry over to the inputs your application actually sees? This talk shows developers how to use imperfect benchmarks to make better decisions about model selection. Using a working document-extraction example, I’ll walk through inspecting benchmark failures, separating model errors from scoring problems, and using semantic assessment to investigate disagreements. We’ll cover the limitations of LLM judges, the importance of preserving application requirements, and what individual examples can reveal that aggregate scores miss. We’ll also explore how fine-tuning to benchmark data can produce strong scores, and what developers should check to assess whether that performance carries over to data outside the benchmark. Attendees will leave with practical ways to assess benchmark claims, investigate whether model performance generalises, and make existing evaluation datasets more useful for choosing and improving models.