Training a Model Over Multiple Epochs

Imagine you have finished setting up a training run. You pass a small collection of text through the model again and again, and after each pass the number that measures how wrong the model is on that same text goes down. It is tempting to read that as success. The training function is doing its job, and the loss it reports is shrinking. This article asks a more careful question: what does a falling training loss actually tell you, and when can it mislead you?

Repetition is normal and useful in training. The concern this article addresses is what happens when the pool of examples is small. Each pass then shows the model nearly the same material as the one before it, which gives the model many chances to learn the particular examples it has been given rather than the broader pattern behind them.

The source explains that the model begins to overfit after a few epochs. The reason it gives is the combination of two things: the dataset is very small, and the training process goes over it multiple times. Trace what the model sees as the epochs accumulate. In the first pass, the examples are new to it. By the next pass, it has already been adjusted on those same examples. By the later passes, the model has met each example many times, and the variety it experiences has not grown.

The diagram shows the progression in words rather than numbers. The source does not give a specific epoch count at which overfitting starts. It says only that it can begin after a few epochs, and that the condition depends on a very small dataset being repeated.

The key idea is the combination. A small dataset alone is not the whole story, and repeated passes alone are not the whole story either. Overfitting appears when both are present: little variety in the data, and many repeated exposures to that same variety. Remove either ingredient and the risk the source describes is reduced.

The source makes a balanced statement about the same run. It notes that the model overfits, and then it adds that the function is working, because it minimizes the training set loss. Both claims are true at once. The training function is correctly driving down the loss it is asked to minimize. The concern is that this particular loss only measures performance on the data the model is already being trained on.

A falling training loss therefore shows that the training function is minimizing the loss it measures. It does not show how the model behaves on text it has not been trained on. A model can keep improving on the examples it has memorized while getting worse on material it has not seen, and the training loss curve would still be heading down.

The source does not stop at diagnosing the small-dataset problem. It invites the reader to train the model on a larger text dataset and to compare the results of this more sophisticated training function with the results from the simpler train_model_simple function. The point of the comparison is to see how the training outcome changes when the data is bigger and the training approach is different, rather than to accept one loss curve at face value.

When you run that comparison, keep one question in view: is the improvement you see a feature of the training function, or a feature of having more varied data? A larger and more varied dataset gives the model less opportunity to memorize a few passages, which is exactly the condition the small-dataset discussion warned about.

The appendix that introduces LoRA sits next to the overfitting discussion for a reason. Both concern how a model is adjusted after its initial training, and both require careful reading of what the adjustment achieves. The source frames LoRA as a fine-tuning technique rather than as a new way of measuring results. It is best understood as a method for changing a model efficiently.

The source describes LoRA as a widely used parameter-efficient fine-tuning technique. Its illustration in the source is the spam classification fine-tuning example from chapter 6. The source also states that LoRA fine-tuning is applicable to the supervised instruction fine-tuning covered in chapter 7. So the technique is not limited to one task; it is tied to these two settings in the source material.