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

A machine learning practitioner documented three training data failures from the past year, each of which produced clean runs with no 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 by learning answer formatting alone. The second failure involved a mismatch between training and evaluation: the model was trained on free-text answers but evaluated on multiple-choice indices, causing loss to collapse near zero while the task metric scored only 0.10. A third issue involved a large public robotics dataset whose usable portion was far smaller than its headline size, buried in the paper's methodology section. The author now treats sudden loss collapse as a warning sign and audits dataset composition and objective alignment before any training run begins.