
A training run is not converging. The curve in front of you narrows the cause to two or three possibilities, and most engineers reach for the learning rate regardless of what the curve says.
Nothing is learning. Check, in this order: is the loss actually connected to the parameters, is the learning rate so small that updates are lost in floating point, are the labels shuffled relative to the inputs, is the model in evaluation mode. The fourth one has cost more collective hours than it has any right to.
The model learned the mean and stopped. This is usually capacity or optimization, not data. Try a longer warmup and a higher peak learning rate before adding parameters.
Typically a bad batch or an outlier gradient. Clip gradients, and log the batch index at each spike; a repeated index means data, not optimization.
The run is dead; the optimizer state is corrupted. Roll back to the last checkpoint, skip the offending batch and lower the learning rate on resume. Continuing costs days.
The textbook overfit, but check for leakage first, because a validation set that shares rows with training produces a curve that looks fine right up until production.
Almost always overflow in an exponential, a log of zero, or a division by a near-zero denominator. Binary-search the graph with anomaly detection on and you will find it in minutes. In mixed precision, suspect the loss scale first.
Write down what you think the curve says before you change anything, then change one thing. A run diagnosed from a hypothesis teaches you something; a run fixed by chance teaches you nothing and you will be back here next month.

Reaching for a neural network on tabular data is a common way to look junior in a system design interview.

Not keyword stuffing, and not a new resume per job. Reordering the truth, with the numbers attached.

Bigger model or more tokens? The question has a numerical answer, and it comes up in interviews more than people expect.