How to evaluate an LLM before deployment

Five questions in a demo prove little

A model answers five questions correctly during a demo. The next day it gets a page scanned upside down, a contract without the table from its annex and a question about a document the person asking has no access to. Only cases like these show whether the system is fit for work.

Before we choose a model, we agree what a correct result is and which error blocks deployment. In a law firm, the answer must point to the right clause and its version. At a help desk, what counts is the department the case goes to. In accounting, a wrong supplier identifier is a different problem from a stray space. The process user helps build the test, because they know which mistakes cause more work.

What it costs and when it pays back

The count covers the time spent searching, reading and compiling material. What remains is the time to check the result against the source, 25% of today's work in the scenario. We measure both times in a trial on 20 to 100 cases the team has already handled.

What stays after the test

We keep the previous working version and the conditions for rolling an update back. We run the same set after every change to the model, the prompt or the data, to compare the result with the previous one. A demo result without such a set does not let you judge whether the new version is better.

Measure several kinds of error

In classification we show the results for each category, because a good average can hide weak handling of rare cases. In extraction we count a missing field and an invented field separately, and in RAG we check whether the cited source supports the answer. A score of 0.9 next to an answer does not prove that nine in ten similar decisions will be correct, so we measure the confidence threshold on data from the task, together with the share of cases passed to a person. The approach is described in OpenAI's evaluation guidelines, and decision models in the article on Jev.

Where to start

The first step is a set of 20 to 100 cases the team has already handled, with the expected result. The report contains the model version, the configuration, the results by category, examples of errors, a load measurement and a recommendation. We deploy local models as part of private LLMs, and we build the view for working with sources as an AI app. For the free review, a description of one case and the places where the materials are kept is enough. Confidential files stay in the company, and you receive the written plan within 2 working days.