Data Labeling QA — Done Right

If you don't trust the people who labeled your training data, do not fine-tune on it as-is. Ten minutes of audit catches errors that will otherwise silently poison your model. Bad labels don't just hurt accuracy — they teach the model the wrong thing, and you won't find out until production.

This skill runs four complementary audits and combines them into a per-row trust score plus a prioritized review set. Each audit catches a different failure mode; running only one leaves blind spots.

When to use this skill