Two few-shot sets, one ambiguous test set

Few-shot examples chosen for matching phrasing and format train the model to mimic surface style. On ambiguous inputs that don't share that phrasing, it falls back to guessing rather than applying a rule.

• Best for: narrow, low-variance tasks
• Watch: fails on inputs with new phrasing

Few-shot examples chosen to sit right at the category boundary force the model to infer the actual rule separating classes. Ambiguous inputs get resolved by that rule instead of by surface similarity.

• Best for: durable transfer across model updates
• Watch: needs careful selection of edge cases