Fine-tuning pitfalls

Fine-tuning has a small number of classic ways to fail, and they are all preventable if you know to look for them. The symptoms are subtle, which is exactly why they catch people.

The big two

1. Catastrophic forgetting. The model gets better at your task and worse at everything else it used to know. This shows up when your fine-tune nails the format but the underlying reasoning got worse.
2. Overfitting. The model memorises your training examples instead of learning the general behaviour. It performs great on data it has seen and falls apart on anything new. This is what the hold-out set exists to catch.

The subtler ones

• Data leakage. Test examples end up in the training set, so your "great results" are just the model reciting what it memorised. Your hold-out set must be genuinely held out.
• Too few examples. A handful of examples teaches the model to parrot, not to generalise. Below a certain volume, fine-tuning is worse than a good prompt.
• Bad labels. If your "good" examples are inconsistent, you are teaching the model to be inconsistently wrong. The quality of the labels is the ceiling of the result.
• Judging on loss, not task. The training loss goes down and the actual answers did not improve. Loss is not your goal; the task is.

How to protect yourself

1. Keep a clean, untouched test set and run it after every training run.
2. Evaluate on the task, not on the training metric: does it actually behave the way you specified?
3. Compare against the untuned base on the same task. If the fine-tune is not clearly better, do not ship it.
4. Watch for forgetting by also testing a few general questions the model used to handle.

The bottom line

Fine-tuning fails quietly. The only defence is measurement: a held-out set, a real task metric, and a comparison to the base model. Without those, you cannot tell the difference between a great fine-tune and a confident mistake.

This is why the method track puts evaluation before everything. Train last, measure first.