Baseline

Ordinary warmup + cosine-decay training on the full dataset. Nothing wrong with it.

0 error0 warning0 info

Divergence

320 healthy steps, then a simulated LR-schedule incident blows the run up.

3 error1 warning2 info

Overfitting

40 training examples, no regularization -- eval loss turns upward after step 125.

0 error1 warning0 info

Missing warmup

Learning rate jumps straight to its peak from step 0, no ramp-up.

0 error3 warning0 info

Throughput drop

A real, measured mid-run slowdown (an actual time.sleep, not a fabricated number).

0 error3 warning0 info

Real transformers.Trainer run

A genuine 2-layer GPT-2 trained via transformers.Trainer, not the numpy MLP.

0 error0 warning0 info