Fine-tuning changes a model’s weights so it performs better on your task than the general purpose version did. That is the whole idea. Everything else — LoRA, adapters, learning rates, epochs — is machinery for doing it efficiently and without breaking the model in the process.
The practical question is rarely “should the weights change” but “how many of them, and by how much”. Full fine-tuning updates every parameter, which needs enough memory to hold the model, its gradients, and optimiser state at once. For an 8B model that is well beyond a single consumer GPU. Low-rank adaptation sidesteps this by freezing the original weights and training a small number of new ones alongside them — typically under 1% of the total. The result is close to full fine-tuning quality on most task-specific work, at a fraction of the memory.
What tends to go wrong
Fine-tuning has a specific and unintuitive failure mode: the model gets worse. Not worse than it could be — worse than the base model you started from. This surprises people, and it usually comes from one of three places.
Overfitting. Training too long on too little data. The model memorises the training examples and stops generalising. Training loss keeps falling while performance on held-out data flattens or degrades — which is exactly why a held-out test set is not optional.
Catastrophic forgetting. Push the weights hard enough toward your task and the model loses capabilities it had before. On a narrow task this can be acceptable. If your application needs the model to also handle general instructions, it is not.
A learning rate that does not match the method. Full fine-tuning and LoRA want different learning rates, often by an order of magnitude. Carrying a value across from one to the other reliably produces a bad run.
None of these are exotic. They account for most of the “we tried fine-tuning and it did not work” stories.
What you actually have to decide
Less than you might expect. Task type and student model matter a great deal. Rank and learning rate matter somewhat, within a fairly forgiving band. Most other knobs matter far less than the quality and diversity of the training data, which is the thing worth spending your attention on.
That is the honest summary: a well-chosen 1.7B student trained on a thousand diverse, validated examples will beat a carefully tuned run on two hundred repetitive ones, and the gap is not close.
Where this cluster goes
The articles below cover LoRA rank and how to set it, the difference between LoRA and QLoRA, what full fine-tuning actually changes that adapters do not, how long training takes, and diagnostics for each of the three failure modes above.
For where the training data comes from in the first place, see the training data cluster. For how to tell whether the result is good enough to ship, see evaluation.