Archive — one question answered
Overfitting: 100% train, 70% test
Train accuracy much higher than test — what is that?
Overfitting means the model memorizes the training set but fails on new data. Example on 40 toy emails: after 500 steps, train accuracy = 100% but test = 70% — a 30-point gap is a warning.
That is not always a bug; often the model is too flexible for the data, or training runs too long without stopping. The course baseline: split test before peeking, compare both numbers, stop when test worsens while train still looks great.
Illustrative table (made-up counts, same split):
| Round | Train acc | Test acc |
|---|---|---|
| 50 | 85% | 82% |
| 500 | 100% | 70% |
When test drops while train rises, log the epoch and shorten training or simplify the model. The track does not teach dropout immediately, but teaches spotting it with numbers — the skill every regularizer builds on.