Three Silent Data Bugs That Corrupted ML Training Without Raising Any Errors

A machine learning practitioner documented three cases where training datasets produced misleading results without triggering any errors or warnings. In the first case, 630 of 688 training records belonged to a single category, leaving several evaluated categories entirely unrepresented, yet the model still showed apparent gains due to format learning alone. In the second case, a near-perfect loss curve and 0.99 token accuracy masked a fundamental mismatch: the model was trained on free-text answers but evaluated on multiple-choice index selection, two entirely different tasks. A third issue involved a large public robotics dataset whose headline storage figures obscured a critical detail buried in the methodology — that only around 60% of the data met usable criteria. The author concludes that silent dataset failures are among the most dangerous bugs in ML workflows, advocating for explicit pre-run audits of data composition and alignment between loss objectives and evaluation metrics.