What Is Fine-Tuning?
Also known as: Model Fine-Tuning, Supervised Fine-Tuning
Definition
Fine-tuning is the process of continuing to train an already pretrained model on a smaller set of your own examples, adjusting its weights so it reliably produces the formats, vocabulary, and behavior your task requires.
Fine-Tuning Explained
Mechanically, fine-tuning runs the same gradient descent used in pretraining but starts from trained weights, uses a much smaller dataset, and applies a far lower learning rate. The model sees your input and output pairs, computes loss against the desired output, and nudges weights to reduce that loss. Because the starting point already encodes language and reasoning, thousands of examples can produce large behavioral shifts where pretraining needed trillions of tokens.
The dominant approach today is parameter-efficient fine-tuning rather than full-weight updates. Techniques such as LoRA freeze the base model and train small adapter matrices instead, cutting the GPU memory requirement by roughly an order of magnitude and producing artifacts measured in megabytes rather than hundreds of gigabytes. That makes it feasible to maintain several task-specific adapters over one shared base model and swap them at serving time.
Data quality dominates data quantity. A thousand carefully reviewed, consistent examples typically outperform ten thousand noisy ones, because the model learns whatever patterns are actually present, including your labeling mistakes. Examples must match production input distribution: if the model will see scanned OCR text with errors, train it on scanned OCR text with errors. Always hold out an evaluation set the model never trains on, and measure against the untuned baseline before declaring success.
Fine-tuning is frequently chosen for the wrong reason. It teaches behavior effectively and teaches facts poorly. If the goal is for a model to know your current inventory, your part catalog, or this week's engineering change orders, retrieval is the correct tool and fine-tuning will produce a model that confidently states outdated information. Fine-tune when you need consistent structured output, a specialized vocabulary, a particular tone, or higher accuracy on a narrow classification task.
Why It Matters
- Produces reliable structured output such as valid JSON for ERP integrations, where a general model drifts in format often enough to break pipelines.
- Teaches domain vocabulary, part nomenclature, and internal shorthand that a general model misreads or silently normalizes away.
- Lets a small fine-tuned model match a much larger general model on a narrow task, cutting serving hardware and latency substantially.
- Bakes knowledge in at training time, so any fact taught this way goes stale and requires retraining to correct.
In Practice
A common gotcha: teams fine-tune on examples written by a subject matter expert in careful prose, then deploy against messy operator-typed input full of abbreviations and typos. Accuracy in testing looks excellent and production performance disappoints, because the training distribution never resembled real traffic.
Frequently Asked Questions
How many examples do I need to fine-tune a model?
For format and tone adjustments, a few hundred high-quality examples often suffice. For narrow classification or extraction tasks, one to five thousand curated examples is a common working range. Beyond roughly ten thousand, returns diminish quickly unless the task is genuinely complex. Consistency and coverage of edge cases matter far more than raw count.
Will fine-tuning make the model forget its general abilities?
Full-weight fine-tuning on a narrow dataset can cause catastrophic forgetting, where general reasoning and instruction-following degrade. Parameter-efficient methods such as LoRA largely avoid this because the base weights stay frozen and the adapter can be removed. Keeping learning rates low and mixing in some general instruction data also helps preserve broad capability.
Related Terms
LoRA (Low-Rank Adaptation)
LoRA (Low-Rank Adaptation) is a fine-tuning method that freezes a model's original weights and trains small low-rank matrices injected alongside them, achieving most of the benefit of full fine-tuning at a small fraction of the memory cost.
RAG (Retrieval-Augmented Generation)
RAG (Retrieval-Augmented Generation) is an architecture that searches your own document store for passages relevant to a question, inserts them into the LLM prompt, and instructs the model to answer only from that retrieved evidence.
Model Weights
Model weights are the numeric parameters inside a neural network, learned during training, that determine how input is transformed into output. They are the model: distributed as files, they encode everything the system has learned.
Go Deeper
GPU Sizing Calculator for LLM Inference
Work out how many GPUs you need to serve a given open-weight model to your user base, based on memory footprint and token throughput.
Small Language Models vs LLMs for Enterprise Workloads
Small language models vs LLMs for enterprise workloads: accuracy, cost, GPU footprint, and when a 7B model beats a frontier model for ERP automation.
On-Prem LLM Deployment Architecture: Reference Guide
Reference architecture for on-prem LLM deployment: inference servers, GPU sizing, RAG pipelines, and security zones for regulated manufacturers.
Working with Fine-Tuning in a live environment? Our engineers do this every day - and our AI agents automate most of it.