Why labels limit the model
A model can only be as consistent as its labels. Wrong labels teach the wrong thing, and inconsistent labels between labellers set a ceiling on accuracy that more data will not lift.
1. Write guidelines with hard examples
For each class, show clear examples and the borderline cases: a partly visible object, a sarcastic message, a scratch versus a reflection. Update the guidelines when labellers disagree.
2. Measure agreement
Have two people label the same sample and compare. Where they disagree, the guideline is unclear or the case is genuinely ambiguous. Both are worth knowing before training.
3. Look for missing and wrong labels
Check for objects without a box, boxes without an object, labels outside the allowed list and text with no label. Review a random sample of each class, not only the ones a model gets wrong.
4. Check balance and coverage
Count examples per class and per setting (camera, location, customer, language). Rare classes and missing settings are where models fail in production.
5. Keep labels and splits clean
Make sure near copies of the same item do not carry different labels, and do not sit on both sides of the training and test split.