Eval refusal classifier
Judges: model response. Questions: choice:type - answered/refused/hedged/off_topic. Action: Count refusal rates per prompt set.
AA clean typed decision (yes/no, a pick from a list, or a level on a scale) that is concrete, repeats routinely, and has meaningful value.3.82
Key facts
- Vertical
- Software & tech
- Function
- Agents & dev
- Status
- Idea
- Volume
- routine
- Value
- meaningful
- Risk
- low
- Evidence
- —
- Flags
- None
Build this with a classifier
Define a typed decision with a bounded answer, then evaluate it on examples.
{
"decision_type": "choice",
"question": "Does this input match the decision in “Eval refusal classifier”?",
"input": "<input to classify>",
"output": "one label from a fixed list"
}Related use cases
Cite this
Copy a link in your preferred format.