You needed to know
whether the new prompt was better,
and there were four hundred cases.
Nobody reads four hundred answers.
So a second model read them.
Pick the better of each pair.
The numbers came back clean.
The new version won,
sixty one to thirty nine,
and it went out on Thursday.
Here is some of what
the judge was measuring.
Length.
Show a model two answers
and the longer one
that says the same thing
will often score higher.
It reads as thorough.
Order.
Put the same pair
the other way round
and a fair share of verdicts flip,
because whichever came first
had a small advantage
nobody asked for.
Tone.
A confident wrong answer
with headings and a summary
beats a hesitant right one
more often than you would like.
And family.
A judge tends to prefer answers
that sound like itself,
so if it shares a lineage
with the model under test,
it is partly marking
its own homework.
None of this is shocking.
Human markers have tastes too.
The difference is
that nobody checked this marker
before handing it the pen.
So the prompt you shipped
may be better.
Or it may simply be longer,
more formatted,
more sure of itself,
and you have tuned your product
toward the taste of a grader
your users do not share.
Treat the judge
like any other instrument.
Calibrate it first.
Take fifty cases,
have people mark them,
and see how often
the judge agrees.
If it cannot match
your people on fifty,
its four hundred mean very little.
Run every pair twice,
swapped,
and count only the verdicts
that survive the swap.
Put the length of each answer
beside its score,
and watch whether one
is following the other.
Use a judge from a different family
when you can.
And keep reading a sample yourself,
ten a week,
picked at random,
not picked by the score.
An automatic grade is cheap,
and it is a measurement.
Measurements have errors,
and this one has a direction.
Find the direction
before it finds its way
into your product.
– Serguey Asael Shinder
Top comments (0)