การทดสอบพฤติกรรมของโมเดล

One accuracy score hides how a model truly behaves. Behavioral tests probe how predictions react when you change the input on purpose.

Behavioral tests check inputs and outputs only, ignoring the model internals. This black-box view works for any framework you serve.

Some changes should not move the prediction at all. An invariance test swaps a neutral word and asserts the output stays the same.

Other changes should move output a known way. A directional test adds a positive word and asserts the score goes up, not down.

Check the easy cases a model must never miss. A minimum functionality test asserts an obvious example gets the obvious label.

These three test types come from the CheckList paper for NLP. The pattern generalizes to tabular, vision, and any model you ship.

Run the same case across groups. A slice test asserts prediction quality stays steady when only a sensitive attribute changes.

Add a typo or jitter a number and the label should hold. A robustness test guards against tiny, harmless perturbations.

Lock the output for a fixed example so retrains do not drift it silently. This regression test flags surprise behavior changes.

Behavioral tests are ordinary pytest functions. They live beside your data tests and run in the same CI step on every push.

Each behavioral test is a written rule the model must obey. Together they form a living specification of expected behavior.

Adding a stronger positive word should raise a sentiment score. Which test is that?

You now run behavioral tests: invariance, directional, and minimum functionality cases that turn expected model behavior into checks CI enforces. ✅