The few-shot examples format inputs as "Input: <text>" and
outputs as "Output: <label>". But the actual query is appended
as "Question: <text>". The model sees a format SHIFT and may
emit unpredictable output — sometimes following the example
format, sometimes following the query format, sometimes mixing.
Fix the query template to use the same "Input:" / "Output:"
pattern as the examples.
Expected output:
Input: late but quality was fine
Output:
The break is on line 5 — but read the whole snippet first.