Supervised Fine-Tuning (SFT)

Supervised fine-tuning (SFT) is further training of a language model that has already been trained, on pairs of inputs and the exact outputs you wish it had produced. A few more passes of training make those outputs more likely, so the behaviour you showed it becomes its default. The architecture does not change. The same weights move toward your examples, and a model that had to be talked into a format on every call starts producing it unprompted.

The short version of when to use it: fine-tuning changes how a model writes, not what it knows. Use it for form, such as a fixed structure, a professional register, or a narrow classification. Do not use it to add facts, and especially not to make citations trustworthy.

How a supervised fine-tuning run works#

You assemble a file of example conversations, each ending in the response you want, and train the model to make those responses more likely. On a hosted platform you upload the file and the provider runs the job. With an open-weight model you run the same loop yourself, almost always through a small adapter rather than updating every weight.

The hosted route has narrowed. As of October 2026, OpenAI's fine-tuning documentation says the company is winding down its fine-tuning platform: new users can no longer reach it, and eligible existing customers can create training jobs until 6 January 2027. A fine-tune is also tied to its base model, so when the provider retires that base, the tuned version goes with it. Plan on open weights you can host if you are starting a fine-tune today, or confirm the provider's retirement dates before building on a hosted one. Otherwise the base model's retirement date becomes your retraining date.

Try the simpler options first#

Each way of adapting a model costs more and commits you further than the one before. A better prompt takes minutes and is reversed instantly. Few-shot prompting, a form of in-context learning, pastes worked examples into the context and fixes many format problems for the price of extra tokens on each call. RAG supplies facts at query time and keeps them editable. Fine-tuning comes last, because you now own a dataset, a training pipeline, an eval suite and a model artifact that has to be versioned, hosted and retrained when the base model or the task changes.

What it fixed on our trademark platform, and what it could not#

Our trademark drafting platform, in production since 2023 and approved by the attorney it was built for, shows both sides.

The drafts had four distinct failure modes: drifting from the required structure, fabricating citations, applying the wrong jurisdiction, and slipping out of legal register. Fine-tuning on the attorney's past opinions taught the model the reasoning pattern, the section boundaries and the register. Of the four, register drift is the only one fine-tuning alone handled well. It does not, and cannot, handle citation integrity: training improves how a citation is phrased, which is the wrong thing to improve when the authority behind it might not exist. Citations there come from retrieval over a whitelisted index of case law, statutes and registry data, structure is enforced by schema-constrained decoding, and a validator rejects any citation that does not resolve.

So fine-tuning was one layer of four, and the layers that made the output trustworthy were the ones outside the weights. If your failures come from missing or outdated facts, the fix is retrieval. If they come from format, tone or a narrow behaviour, fine-tuning is worth trying.

The dataset matters more than its size#

The model imitates whatever you show it, including inconsistent labels, uneven formatting and contradictions. On the trademark build, normalising the corpus came first and mattered more than we expected. Years of attorney opinions were unified into one five-section template and tagged by mark type, jurisdiction and NICE class. Without that, the model would have learned a dozen individual writing styles and averaged them.

On size, OpenAI's guide sets a minimum of 10 examples and reports improvements from 50 to 100 well-written ones. Treat that as where a first run starts, not a target. Remove duplicates, fix disagreements between labels, make the formatting uniform, and hold out a test set the model never trains on, with evals written before you train so the run has to show an improvement against them.

Adapters, forgetting and distillation#

Updating every weight is expensive and produces a whole new copy of the model. The usual route is LoRA: freeze the base model and train small low-rank matrices beside it. On GPT-3 175B, the LoRA paper cut trainable parameters by 10,000 times and GPU memory by 3 times, with no added inference latency. Adapters are small files, so you can keep one per task or per customer. QLoRA goes further by storing the frozen base in 4 bits, and its paper fine-tuned a 65B model on a single 48GB GPU.

The standing risk is catastrophic forgetting: train hard on a narrow task and the model can lose general skills, such as following instructions or refusing what it should refuse. A low learning rate, few epochs, some general data in the mix and a frozen base all limit the damage. Re-run broad evals after tuning, not only the task evals.

Distillation is the related move for serving cost: have a large model produce outputs for your task, then fine-tune a small model on them. The idea goes back to Hinton and colleagues in 2015.

Keep the fine-tune replaceable#

A fine-tune ages. The base model improves or is retired, and your tuned copy does not follow. On the trademark platform, production logging scores the base model and the fine-tuned model on the same inputs with a 0 to 10 legal-integrity evaluation across six criteria, so a regression from a model change shows up as a score that moved. The attorney's corrections stay the most valuable signal, because they measure what automated scoring cannot see.

Bigcircle recommends that comparison for any fine-tune, because it shows whether the tuned model still beats a newer base model given a good prompt. When it stops winning, retire it.