JwumaModel evaluation, domain experts versus generalists working from a checklist
· 2 min read · evaluation, expert data
Generalists can reliably evaluate tone, helpfulness, format and obvious safety failures. They cannot evaluate whether a clinical or legal answer is correct, because that requires knowing the domain. Commissioning generalist evaluation for expert content produces confident scores that measure fluency rather than accuracy.
The short version
- Generalists evaluate surface qualities reliably: tone, format, helpfulness.
- Domain correctness requires an evaluator qualified in that domain.
- Fluent wrong answers score well with unqualified evaluators.
- Failure taxonomies are more actionable than aggregate scores.
- Evaluation in a language should be done by speakers of that language.
Scoping something like this?
We would rather talk about the problem than quote against a volume figure.
See capabilitiesWhat can a generalist reliably judge?
Whether an answer is clear, appropriately hedged, correctly formatted, responsive to what was asked and free of obvious safety problems. These are real and worth measuring, and a well-calibrated generalist panel does them well and cheaply.
What they cannot do is tell you whether the answer is right in a domain they do not know. That is not a training problem, it is the definition of the domain.
What goes wrong when you get this backwards?
You get high scores that measure the wrong thing. A fluent, confident, well-structured and completely incorrect clinical answer scores well with an evaluator who cannot assess clinical correctness, and the resulting metric actively misleads you about model safety.
This is worse than having no evaluation, because a number gets reported and acted on.
How do you tell which you need?
Ask what a wrong answer would look like. If a wrong answer is recognisable to an intelligent layperson, generalists are fine. If a wrong answer looks exactly like a right answer unless you know the field, you need a practitioner.
Most evaluation programmes need both, on different axes of the same output, rather than choosing one.
- Would a layperson recognise a wrong answer?
- Does the failure mode look like fluent competence?
- Is the content regulated or safety-critical?
- Is the output in a language your evaluators actually speak?
What should the output be?
A failure taxonomy, not just a score. Knowing a model scores 72 tells you nothing actionable. Knowing that its failures cluster into three named modes, one of which is safety-relevant, tells you what to fix.
Ask for the taxonomy as a deliverable at specification time. Retrofitting one from scored data afterwards rarely works.
People also ask
When do I need domain experts for model evaluation?
When a wrong answer would look like a right answer to a non-specialist. If the failure mode is fluent competence, only someone qualified in the domain can detect it.
Can generalist evaluators assess clinical or legal output?
They can assess tone, format and obvious safety issues, but not correctness. Using them for correctness produces scores that measure fluency and mislead about model safety.
What should an evaluation programme deliver besides scores?
A failure taxonomy naming the distinct ways the model fails. A score alone is not actionable; named failure modes are.