13 Fine-tuning

Fine-tuningⓘfine-tuningtrainingContinued training of a pretrained base model on a smaller, task-specific dataset to specialize its behavior without retraining from scratch. Chat about this → Open full entry → changes model behavior by training on additional data. For local users, the two methods that matter most are LoRAⓘLoRAtrainingA parameter-efficient fine-tuning method that injects small low-rank adapter matrices into a frozen base model, training a tiny fraction of weights instead of the full model. Chat about this → Open full entry → and QLoRAⓘQLoRAtrainingA fine-tuning method that combines 4-bit quantization of the frozen base model with LoRA adapters, making large-model fine-tuning fit on a single consumer GPU. Chat about this → Open full entry →. LoRA freezes the base model and trains small low-rank adapter weights, which reduces trainable parameters and lets you keep multiple lightweight adapters. QLoRA extends this by fine-tuning through a frozen 4-bit quantized model into LoRA adapters, which is what makes fine-tuning feasible on a single consumer GPU.

Fine-tune when you need a consistent writing style, a domain-specific output format, repetitive classification or extraction behavior, tool-call format reliability, a specialized persona, domain adaptation that retrieval cannot solve, or better small-model performance on a narrow task.

But do not fine-tune first. Try this order: correct chat template, better prompting, a better model, better decoding, RAGⓘRAGretrieval-memoryA pattern where a model retrieves relevant documents from an external store at query time and conditions its answer on them, instead of relying only on parametric knowledge. Chat about this → Open full entry →, reranking, few-shot examples, and only then fine-tuning. Most problems that look like “the model does not understand my domain” are actually a vague prompt, a wrong template, or broken retrieval. Fine-tuning is the most expensive lever and the slowest to iterate, so it should be the last one you reach for, not the first.

When you do fine-tune, a sound plan includes clean data, train/validation/test splits, baseline evals, a clear target behavior, a safety review, overfitting and regression checks, adapter versioning, license review, and a rollback plan. The tooling for pretraining and fine-tuning lives at the training layer.