Good instructions have good examples. So do good codebooks.
Definitions define the space. Examples populate it.
##intervention_type
Question
What was the main intervention?
Description
The therapy, medication, or behavioural program administered to the treatment arm.
Example answers
"Cognitive Behavioural Therapy (CBT), 12 weekly 60-minute sessions, delivered in group format."
Default if missing"Not specified"
An example can carry more, or less.
●○○○one ordinary answerwhat a filled cell looks like, shape included
●●○○a key source quote and the answerthe mapping, not just the output
●●●○and the reasonwhich feature of the document decided it
●●●●two that look alike, answered differentlywhere the boundary is
@karlrohe
Examples are old technology, and they work on AI readers for the reason they work on people. When you are learning to answer a question, you want to see one answered. You get to see how the rule is actually used. Nothing about that is new, and nothing about it is specific to AI.
A prompt with no examples has a name: zero-shot. Put a few in and it is called few-shot. Computer scientists love to rename everything.
definitions
zero-shot — instructions with no worked examples attached. Every codebook field starts here.
few-shot — the same instructions with a handful of examples: if the document said this, the answer is that.
Examples come out of reading you are already doing. Read a document, hit a case where you have to stop and decide, and you have just made an example. Write down what the document said and what you chose. Keep an eye out for the good ones as you go, and put them in.
How far you go depends on what went wrong. A cell that comes back in the wrong shape needs one ordinary answer. A cell where readers split needs the reason with it. A boundary that keeps moving needs the pair that straddles it.
List and table fields always carry a format example. The reader has to be shown the JSON shape it is expected to return. That example is a format contract, not a reading of your intent.
discussion
The field on this card is a real one, and it is not a very good question. Why not?
hint
You cannot tell from the question alone. What would you need to know about the corpus first? And what happens when a trial has more than one?
one answer
Three cases the question cannot hold. There may be several interventions, and they may overlap, which wants a list rather than a string. Arms and interventions may mix and match, and then they are two lists joined together, not one cell. The study may have no intervention at all, and in an observational corpus the word is simply wrong.
In each case the instructions do not say what to do. That is the defect. Not that the documents are complicated, but that the codebook is silent about them. You find the silences the way you find everything else: the readers disagree, and the disagreement points back at the question.