training data, prediction, bias, and accountability
Training shift
A crop-disease model trained on bright laboratory photos scores 94% validation accuracy. On farmers' low-light phone photos, accuracy is 61%.
- Validation matched training conditions.
- Deployment distribution differs.
- Collect representative data and test by condition.
What should the team do next?
MODELModel Responsibility Card
DATA
PROVEChoose with evidence
CLUEPause. Find what matters. Then make your move.