The definitive guide
to fine-tuning.

Every generation hands its craft to the next the same way. Learn the general ideas, then learn the trade beside someone who corrects you.

That is how a guild trained a silversmith, how a hospital trains a surgeon and how an airline trains a pilot. Nobody found a faster way. The next slides show the pattern in three places, then show the day it was written into software.

Before the residency, many doctors learned from lectures and then went straight into practice.

William Halsted, who led surgery at Johns Hopkins, helped change that. New doctors stayed on for years and worked real cases beside a senior surgeon. Each year the cases got harder and the supervision lighter. It was the guild method, rebuilt for medicine, and it is still how surgeons are trained.

A report on medical schools put the classroom first and the bedside second.

Abraham Flexner toured the medical schools of the United States and Canada for the Carnegie Foundation and found many of them weak. His fix copied Johns Hopkins. Two years of laboratory science, then two years of clinical work in a teaching hospital. General first, then specific. Language models are trained in the same order.

Nobody has ever made an expert by handing them a manual.

Not a surgeon, not a pilot, not an underwriter, not a silversmith. In every field where getting it wrong matters, the method is the same. Teach the general case in a classroom, then put the person beside real work, let them try, and have somebody experienced correct them. We have known this for eight hundred years. Then artificial intelligence arrived and we started writing manuals again, and calling them prompts.

• Stage one is the classroom. A very large model reads an enormous amount of general material and learns to read, write and reason.
• Stage two is the apprenticeship. The same model is shown examples of the specific behavior wanted, and corrected until it produces it.
• That second stage is fine-tuning. The laboratories run it on their own models before selling them to you, because without it the model is a brilliant graduate who has never held a job.
• So the question was never whether fine-tuning works. It is why the apprenticeship stops at the point where the model meets your work.

Image networks showed that early layers learn what is general and late layers learn what is specific.

A 2014 study trained pairs of image networks and measured, layer by layer, how well each one carried over to a new task. The first layers learned simple shapes and colors that worked almost anywhere. Only the last layers had to be retrained. It was the Flexner recipe, found inside a network.

With only 100 labeled examples, a fine-tuned model matched training from scratch on 100 times more data

The same method cut error by 18 to 24 percent on most of six text tasks. The trick was to pretrain once on general text and then tune briefly on the task at hand.

A hundred examples were enough to start

In 2018 language models learned to be pretrained once and tuned many times.

ULMFiT showed it in January, and OpenAI’s GPT and Google’s BERT followed within the year. Pretrain on general text, then tune on the task. That recipe became the default for language, and it is why tuning was cheap enough to try.

Then GPT-3 sent almost everyone down a different road.

A 175 billion parameter model could do new tasks from a few examples typed into the prompt, with no training at all. Prompting felt like magic, and tuning a model that size was out of reach for nearly everyone. Tuning did not stop working. It stopped being the first thing anyone tried.

Fewer trainable parameters than fully tuning GPT-3 175B with Adam, and 3 times less GPU memory

The paper reports quality on par with or better than full fine-tuning on RoBERTa, DeBERTa, GPT-2 and GPT-3, with no added inference delay.

Freeze the model and train a small add-on

A model with 100 times fewer parameters won, because it was tuned on what people wanted.

In OpenAI’s human evaluations, outputs from the 1.3 billion parameter InstructGPT were preferred to those of the 175 billion parameter GPT-3. The 175 billion version of InstructGPT was preferred 85 percent of the time. Making a model bigger did not make it follow instructions. Tuning did.

The whole cost of Alpaca 7B, under $500 for the training data and under $100 for three hours on eight GPUs

Stanford tuned Meta’s LLaMA 7B on 52,000 instructions. The model hallucinated more than the one it imitated, which is the honest catch. The point was the price.

The data cost more than the compute

QLoRA put a 65 billion parameter model on a single GPU.

The paper reports memory for tuning a 65B model falling from over 780GB to under 48GB with no loss against full 16-bit tuning. Its best model reached 99.3 percent of ChatGPT on the Vicuna benchmark after 24 hours on one GPU. The authors also published where it fails.

Distillation turned one large model into a family of small ones.

DeepSeek tuned open models from 1.5B to 70B parameters on 800,000 samples written by its large reasoning model. Its report says distilling from the large model beat applying reinforcement learning to a small one, and its 32B distilled model beat OpenAI’s o1-mini on several benchmarks.

A prompt is an instruction sheet you hand over again every single morning.

Imagine hiring a capable graduate, and every day before they start you give them the same two-page brief on how your company words things, which exceptions matter and who signs what. They read it, do the work well, go home, and forget all of it. Tomorrow you hand them the sheet again, and you pay for the reading every time. That is prompting. It is an entirely reasonable way to start, and a strange way to run something for three years.

We know more than we can tell. The most valuable things an expert knows are the ones they cannot write down.

A model improves the way a person does, by doing the task and being corrected.

Ericsson found that hours alone make nobody better. The hours that count sit at the edge of your ability, with correction arriving at once. A fine-tuning run is that loop at machine speed. The model attempts a case, sees the right answer, adjusts, and attempts the next one. Thousands of worked examples are thousands of corrected attempts.

Hours alone make nobody better. Hours spent at the edge of your ability, with correction arriving immediately, do.

The things people find hard turn out to be easy for machines. The things any adult does without thinking turn out to be hard.

Most steps in an agent are narrow and repeated, which is where small models belong.

A June 2025 position paper from NVIDIA researchers argues that small language models are sufficiently powerful, more suitable and more economical for many calls inside agents, and that mixed systems with a large model for the hard steps are the natural design. It is a position paper, so read it as an argument and check it against your own measurements.

General methods that use more computation have beaten methods built on human knowledge, again and again.

Renting the frontier is cheap on the first call and costs the same on every call after.

Tuning costs more up front, once, and then each call is cheaper by a large factor. The question is how many calls you make. Below some volume renting wins. Above it, owning does, and the gap keeps widening. The more specific your need, the sooner the line is crossed.

The more specific an asset is to your own use, the worse it is to rent it and the better it is to own it.

The knowledge that matters most is scattered among the people who do the work, and no central body can collect it.

Each producer should do the work it is relatively best at, even if another producer is better at all of it.

A simpler product that is good enough for the low end of a market improves until it serves the middle.

Pre-training, fine-tuning and post-training do three different jobs.

Pre-training builds the base model on trillions of tokens of general text. Model makers do it once, and customers never do. Fine-tuning teaches your domain, vocabulary and formats from question and answer pairs. Post-training shapes tone and judgment using preferences and rewards. You start from an open-weight model every time.

Rented compute for the final run of a 27B model, at $2.50 to $4 per H100 hour, before experiments, failed runs and the research team

That is about four million H100 hours, or roughly six months on a thousand GPUs. Experiments and staff typically double or triple the total. Qwen does not publish its own cost, so this is Lyzr’s estimate. It is why every customer starts from an open model.

Nobody here pre-trains

LoRA wins for almost every customer.

It gets close to full fine-tuning quality on domain tasks at a fraction of the memory, and one base model can carry an adapter for every use case. Use QLoRA when you have a single workstation card, and save full fine-tuning for very large datasets. A 2025 study by Thinking Machines found LoRA matched full tuning on small to medium datasets. It falls behind only when the data exceeds the adapter’s capacity.

DPO teaches a model which of two good answers to prefer.

The model sees a preferred answer and a rejected one to the same prompt and learns to lean toward the preferred one, with no separate reward model. In the cheat sheet’s insurer example both answers are accurate, and DPO teaches the model to prefer the one with empathy, a reason and a next step. It is for taste, tone and policy, after fine-tuning has taught the content.

GRPO teaches reasoning, by scoring several answers and rewarding the ones above average.

The model writes several answers to each prompt. A reward function scores them, and the model learns from the ones that beat the group average. No written answers are needed, only prompts and a way to check. It works best where the answer can be verified automatically, as in math, code, extraction and classification.

Forecast worldwide spending on AI in 2026, up 49.5 percent from $1.79 trillion in 2025

Gartner expects $3.6 trillion in 2027. It says enterprises have yet to flex their spending, and expects 2026 to be the inflection year. It also places generative AI in the trough of disillusionment this year, which is when buyers start counting costs.

The bill is about to get a lot of attention

About 1 percent of that spending is for the models themselves.

Gartner puts spending on generative AI models at $28 billion in 2026, against about $1.5 trillion for infrastructure and $576 billion for services. The intelligence is a thin slice of a very large stack, so how you run a model matters nearly as much as which one you pick.

What enterprises spent on generative AI in 2025, up from $11.5 billion in 2024

That is 3.2 times in twelve months, and $1.7 billion in 2023. In the same report Menlo calls fine-tuning niche and used mostly by frontier teams, which is the gap this guide is about.

From $1.7 billion to $37 billion in two years

Share of production models that were fine-tuned, in Menlo’s 2024 survey

Retrieval led by comparison, at 51% adoption, up from 31% a year earlier. Menlo called fine-tuning surprisingly rare. That is a small base, which is both the opportunity and the warning.

Rare in 2024 and still niche in 2025

Spread between the cheapest and most expensive tier, per million input tokens, on one provider’s public price list

Claude Haiku 4.5 is listed at $1 per million input and $5 per million output. Opus 5.5 is $4 and $20. Fable 5.1 is $10 and $50. Most production agent traffic runs near the top by default, because that is the model the pilot was built on and nobody went back to it.

Nobody chose the expensive tier. They inherited it

Of the three largest providers now offer no first-party way to fine-tune their frontier models

OpenAI is winding down its self-serve fine-tuning service. New customers were shut out in May 2026, and from January 6, 2027 no customer can start a new job. Anthropic has not offered fine-tuning through its own API. Whatever the reasons, the practical effect is the same. If you want a model shaped around your work, it will be an open-weight model, and it will run somewhere you choose.

The decision has been made for you

Of specialized tasks where fine-tuned adapters outperformed GPT-4, across more than 700 experiments

That figure comes from Predibase’s own research on their own platform, so read it as a vendor result and not an independent one. It is still the most quoted number in this field, and it matches what teams report once they stop testing on public benchmarks and start testing on their own cases.

Narrow tasks. Not everything

Gartner’s forecast of usage volume for small, task-specific models against general-purpose LLMs by 2027

Gartner’s reasoning is that general models lose accuracy on tasks that need business context, and that smaller models answer faster and cost less to run. Forecasts miss, so weigh this one as a forecast.

A forecast, so weigh it as one

On a routine task you are paying for capability you will never use.

A frontier model carries the ability to write a sonnet, review a legal contract and explain general relativity. Your duplicate-invoice check uses none of it, and you pay for all of it on every call. In production, most agent traffic is the same handful of steps repeated, such as extract, classify, route and summarize. That is the part worth owning.

Renting starts at nearly zero and rises with every call, forever. Owning costs something up front and then flattens, because the model runs on hardware you already pay for. The crossing point is arithmetic and not opinion, and for a task running thousands of times a day it tends to arrive sooner than people expect.

Everything to the right of the crossing is margin you are handing over

Data preparation decides the result more than the choice of method does.

Clean the raw documents. Drop duplicates, filter junk and boilerplate, remove or mask personal information, turn PDFs, Word files and web pages into clean text, and cut them into passages the model can learn from. Customer documents are plenty for fine-tuning and nowhere near enough for pre-training.

Two hundred documents is plenty to fine-tune and nowhere near enough to pre-train.

Two hundred documents come to roughly one to five million tokens, and eighty thousand pages to about 55 million. Pre-training a 27 billion parameter model needs about 540 billion tokens at the textbook minimum, and real models of that size saw far more. Fine-tuning adapts a base model that already knows the language. It does not need an ocean.

Turn each passage into two kinds of training pair.

A question and answer pair teaches the model what to know. A preference pair, one prompt with a better and a worse answer, teaches it how to answer. Usually a strong model drafts the pairs and subject matter experts review a sample. A few thousand clean pairs often produce a visible shift, and quality beats volume.

The evidence you collected is the evidence that survived. The cases that matter most are the ones that never made it into your data.

Train an adapter with LoRA, shape it with DPO, and expect compute to be the small bill.

On a typical job of 10,000 pairs for three epochs, Lyzr’s cheat sheet puts a hosted LoRA run near $32 on a 27 billion parameter model. Several candidate runs plus DPO and evals keep compute in the hundreds to low thousands of dollars. Data preparation, synthesis and expert review cost more.

Nothing ships until it beats the tests you wrote first.

Test the tuned model on held-out customer data that was never used in training, and against a frontier model baseline. When source documents change or scores drop, re-run data preparation and tune again until a model clears the gate.

A model can improve on the average while getting worse at every single kind of case inside it.

Train in the cloud, download the weights, and run inference on hardware you own.

Training is a short burst that needs fast linked GPUs, while inference runs every day, so inference is the hardware worth owning. Confirm in writing that the provider lets you download the weights and does not keep your training data. Fireworks and Together both offer private cloud options.

Fine-tuning needs examples of the task done well. In 2023 that meant a labeling project, with people, months, a budget line and an argument about quality. Most teams correctly decided it was not worth it.

How the 5% steer clear

• Run the workflow with a frontier model first
• Keep the traces, including the human corrections
• Let a judge model filter the good ones
• Train on what survives

Train a model hard on one task and it gets worse at others. Teams would fix the extraction and find a fortnight later that the summaries had quietly degraded.

How the 5% steer clear

• Train an adapter, not the whole model
• One adapter per task, not one model for everything
• Keep a general test set and run it every time
• Keep the original model to fall back to

Teach a network something new without care and it does not add the skill. It overwrites the old one.

A model trained on last quarter’s forms keeps applying last quarter’s rules. Teams pictured a retraining project every time the business changed.

How the 5% steer clear

• Retrain on a schedule, not on a crisis
• Keep the big model for anything the small one is unsure about
• Watch how often the two disagree
• Facts that change daily belong in retrieval, not in the model

Swapping the model behind a working process is a frightening thing to sign off, and the person signing had nothing to look at but a benchmark score from somebody’s blog post.

How the 5% steer clear

• Run both on real traffic, side by side
• Compare on your cases, never on public benchmarks
• Agree the bar before you start
• Switch one task at a time, keep the fallback

In 2023 this meant a research engineer, a GPU cluster and a tolerance for failed runs. Almost no enterprise outside the technology sector had that.

How the 5% steer clear

• Treat it as a deployment, not a research project
• Start from an open model you are allowed to host
• Use the examples you already have
• Keep the whole loop inside your own environment

Prompt a model directly, and you hand over the instructions again on every call.

You write the instructions and examples into the request. The model is unchanged and you can start in an afternoon, so it is the right place to begin. The cost is that every instruction, example and rule travels with every call, forever, and it only goes as far as what can be written down.

Retrieval plus a model plus tools makes an agent.

Retrieval lets the model look things up in your documents at answer time. Tools let it act, in your CRM, your tickets or your database. Wrapped in a loop, that is an agent. It changes what the model can see and do. It does not change how the model behaves by default.

Fine-tuning changes the model, so the behavior comes built in.

You train the model on worked examples of the job. The format, the tone and the judgment calls become its default, and the instructions you no longer send are the saving. A tuned model holds facts less reliably than a lookup does, so facts that change still belong in retrieval.

The three stack, and a good system uses more than one.

A tuned model can sit inside an agent. Retrieval keeps its facts current, tools let it act, and the prompt carries today’s request. The tuning gives the agent a small, fast, consistent core in place of a large, expensive, variable one. The useful question is which layer each job belongs in.

Prompt first. Retrieve second. Tune what is left.

Most problems are solved by a better prompt, and that costs an afternoon. Of what remains, many are solved by giving the model the right documents. A few stay hard, and they share a shape. They are narrow, repeated, and need behavior that no instruction produces reliably. Tune those, and working in this order keeps the tuning budget for the work that deserves it.

A few items account for most of the volume, and a long tail of rare items accounts for the rest.

The cost saving is the part finance likes. Sovereignty is the part the regulator asks about.

A model you own runs where you tell it to, inside your own cloud account, on your own hardware, or on a machine with no network connection at all. Nothing leaves unless you send it. If a provider trains your model, the weights are the asset you paid for, so confirm in writing that you can download them and that your training data is not kept.

Subject matter experts, not engineers, decide whether the pairs are right.

A strong model can draft thousands of training pairs, but someone who knows the work has to review a sample. That is a few hours a week from your best claims assessor or underwriter, and it is the part nobody can outsource. The training step is a configuration file. The judgment about what good looks like is yours.

A flow is money leaving every month. A stock is something you still have at the end of it. They do not belong on the same line of a budget.

A durable advantage comes from a resource that others cannot get on the same terms.

Newer models follow instructions so well that prompts and retrieval may cover most of the work.

After OpenAI began winding down fine-tuning, a common reading was exactly this, and it is a serious one. Base models improve every few months, and each improvement can erase a tuned model’s edge. Menlo found tuning rare in production. Our answer is narrow. Tune only where the work is repeated, the knowledge is yours and the volume is high, keep the training cheap enough to repeat, and re-test every time the base model changes.

The challenge

Moving one task off a rented frontier model used to mean a research project. The hard parts were never the training run. They were getting clean examples, proving the result and keeping it inside your environment.

What was built

Lyzr’s Open FineTuner connects to your folders, such as S3 and SharePoint, and turns the documents into training data inside your environment. It trains with LoRA then DPO, runs continuous evals, and serves a model to your agents only after it passes.

A change to your documents restarts data preparation, and a failed eval restarts training.

Openfinetuner watches the source folders, so a changed document triggers the data pipeline again and the dataset stays in sync with the business. Continuous evals run on a schedule against a fixed set of test cases. A failure starts refinetuning on the latest data, re-runs the evals, and ships only a model that clears the benchmarks.

It runs where your documents are, so the dataset never has to leave.

Openfinetuner converts documents into training data inside your environment. Pair it with Optimus, Lyzr’s on-premises appliance, and the trained model can run with no external calls. Lyzr lists Optimus as shipping in Q4 2026 with pre-orders open. The eight-card configuration in the cheat sheet, 768GB of memory across eight RTX PRO 6000 cards, is the Optimus Mini.

Lyzr frames ownership as four steps, and fine-tuning is the third.

Lyzr’s Sovereign AI page lays out a progression. Govern the agents you already have from one control plane. Ship reliably with its production modules. Own your models by fine-tuning open source models on your own data. Own the hardware with an on-premises appliance. Lyzr describes it as a progression, with each step added to the last and no rebuild.

Before an agent reaches production, Lyzr runs it through simulations.

Nitro is listed as live. Lyzr says it runs 50,000 simulations before any agent touches production, and it includes SuperFlow for hybrid process orchestration, ShadowLM and agent CI/CD. For a tuned model this is the evaluation gate. Lyzr’s own figures, such as 95 percent savings against frontier API costs, are its claim and have not been independently checked.

Founded in New York in 2023. Agent Studio, Architect, Opencontroller, Open Cowork, Openfinetuner, Optimus and Madison cover building, governing, running and owning agents. Deploys in your AWS, Azure or GCP account, a dedicated tenant, SaaS, on bare metal, fully air-gapped, or on the Optimus appliance.

Backed by Rocketship.vc, BGV, Accenture Ventures, GFT, the Partnership Fund for New York City and Plug and Play.