Data audits and training diagnosis
Repair an evaluation pipeline and diagnose a training failure, with separate project assessments.
What you’ll explore
- Data quality
- Leakage
- Training diagnosis
- Independent transfer
Your lessons
0 / 10 completeWhat should the machine predict?
Define inputs, outputs, users and success before choosing a model. Test how a changed target changes the labels.
Examples are measurements
Inspect missing values, copied requests, unit mismatches and conflicting labels before treating a table as evidence.
Decide where AI belongs
Compare a rule, supplied candidate predictions and simulated human review, including costs and unfinished work.
Repair data without leaking answers
Audit real flower measurements, separate related rows, and turn a tempting score into a defensible claim.
Repair a broken training run
Inspect predictions, gradients and actual parameter changes. Compare a failure and repair from the same checkpoint.
Different mistakes have different costs
Move a threshold, count different errors, and defend a decision using precision, recall and explicit costs.
How much should a small result convince you?
Draw repeated samples, inspect uncertainty and discover why more biased data can strengthen the wrong claim.
Who receives the errors?
Trace a limited review budget through two channels and compare who is missed, reviewed unnecessarily or absent from the evidence.
Use data with a purpose
Choose justified fields, examine provenance and permission, and build a practice data card with explicit handling limits.
Explain what your model is for
Turn results into bounded claims, identify missing evidence and draft a model card linked to your actual experiment.