LLM fine-tuning is additional training of a ready-made language model on a company’s examples so it can consistently answer in the required format, style, and terminology. It adds little or no new company knowledge: for answers based on documents, RAG is usually needed. You can start with a few dozen high-quality examples—OpenAI recommends 50—and most of the work is preparing gold-standard data and checking quality.
Prepared by AIrassvet October 19, 2026. It is intended for executives, CIOs, and product owners deciding whether a model should be fine-tuned for their use case. The answers are based on OpenAI and Yandex AI Studio documentation, academic papers on LoRA, QLoRA, and fine-tuning vs. RAG, open-model licenses, and the text of Russian Federal Law No. 152-FZ on personal data. We did not conduct our own fine-tuning experiments for this article; platform status was verified as of publication and may change.
Contents
- Quick overview
- What this article covers and what it does not
- Questions about the purpose and choice of method
- Questions about data and legal issues
- Questions about cost, timelines, and hardware
- Questions about platforms and models
- Questions about quality and operations
- What fine-tuning gives you and what it does not: summary table
- Primary sources and verification method
Quick overview
- Fine-tuning changes a model’s behavior, not its knowledge. Yandex AI Studio states outright that fine-tuning does not add new information to the model, such as a support knowledge base.
- The entry barrier is low in terms of data, but high in terms of quality. OpenAI’s minimum is 10 examples, with a starting point of 50 well-prepared dialogs.
- LoRA makes fine-tuning accessible. The method’s authors report a 10,000x reduction in trainable parameters and 3x lower GPU memory use for GPT-3 175B.
- Platforms change. OpenAI is phasing out standalone fine-tuning: new tasks for existing customers close on January 6, 2027.
- Personal data in the dataset is a separate legal issue under Federal Law No. 152-FZ, and it must be addressed before collecting examples.
What this article covers and what it does not
This section covers questions about fine-tuning text language models on a company’s examples: supervised fine-tuning (training on pairs of “input — correct answer”) and its lighter-weight variant, LoRA. We do not cover training a model from scratch, fine-tuning computer vision or image-generation models, or a detailed choice between prompts, RAG, and fine-tuning—there is a separate article on that topic: “RAG, fine-tuning, or prompt engineering: which should you choose”.
Terms. *LLM* — large language model. *Base model* — the model being fine-tuned. *Dataset* — a set of training examples. *LoRA* (Low-Rank Adaptation) — a method in which the base model’s weights are frozen and only small additional adapter matrices are trained. *RAG* — an approach in which the model receives retrieved document excerpts before generating an answer.
Questions about the purpose and choice of method
What is LLM fine-tuning in simple terms?
LLM fine-tuning is continuing to train a ready-made model on your examples of “question — reference answer.” The model is not built from scratch: you take a base model, such as open Qwen or cloud-based YandexGPT Lite, and show it hundreds of correct examples. After that, it more often responds in the desired structure, tone, and terminology. The result is either a copy of the model or a small adapter for it that you use instead of the original version.
Will fine-tuning give the model knowledge about our company?
Almost not, and you should not count on it. Yandex AI Studio documentation explicitly says fine-tuning does not add new information such as a support knowledge base. A 2023 study by Ovadia et al. showed that RAG outperformed unsupervised fine-tuning both on facts the model already knew and on new ones. If employees need answers about policies and pricing, start with document search, not fine-tuning.
When is fine-tuning justified, and when are prompts or RAG enough?
Fine-tuning is justified when the task is repeated thousands of times and the model must maintain a strict format: classify tickets, extract fields from emails, or write responses in a branded style. If instructions and a few prompt examples solve the problem, it is too early to fine-tune. If the model lacks facts from documents, you need RAG. First measure errors on a set of real queries, then choose the method based on the type of error.
How is LoRA different from full fine-tuning?
With full fine-tuning, all model weights change; with LoRA, only small additional matrices are trained while the original weights remain frozen. The authors of LoRA (Hu et al., 2021) report that for GPT-3 175B the method reduces the number of trainable parameters by 10,000x and GPU memory requirements by 3x compared with full fine-tuning. For business, that means less hardware and the ability to keep several adapters for different tasks.
Questions about data and legal issues
How many examples are needed for fine-tuning?
OpenAI’s technical minimum is 10 examples, but the documentation recommends starting with 50 carefully prepared dialogs and evaluating the result; noticeable improvements there are associated with roughly 50–100 examples. For classification and field extraction, you usually need more to cover all classes and rare cases. More important than volume is a consistent answer standard: contradictory reference answers will teach the model contradictions.
What format should the dataset use?
Usually it is a JSONL file: one line is one example in a dialog format with system, user, and assistant roles. This is required by OpenAI, and similar formats are used by Yandex AI Studio and libraries for open models. Put the real input and the answer you are ready to sign off as the reference into each example. Keep a separate validation set that is not used for training—without it, you cannot honestly compare the model before and after.
Can you fine-tune a model on customers’ personal data?
Only with a lawful basis and after legal review; it is safer to anonymize examples before training. Under Part 3, Article 6 of Federal Law No. 152-FZ, outsourcing data processing to a contractor requires a contract specifying the data, actions, and purposes. Article 12 requires notifying Roskomnadzor before cross-border transfer, and Part 5 of Article 18 prohibits keeping databases with Russian citizens’ data abroad except in specified cases. This is not legal advice—make the decision with a lawyer.
Who owns the fine-tuned model and the dataset?
This is determined by the contract and the base model license. Under Article 1296 of the Civil Code of the Russian Federation, the exclusive right to a program or database created under contract belongs to the customer unless the contract says otherwise. But the base model weights remain under its license: for example, the Llama 3.1 license requires the name of any distributed derivative model to begin with the word “Llama.” Specify in the contract who owns the dataset, adapter, and training scripts.
Questions about cost, timelines, and hardware
How much does LLM fine-tuning cost?
There is no universal price: the cost includes data preparation, compute, quality evaluation, and ongoing model maintenance. With LoRA, compute is not the only—and not necessarily the main—expense: experts prepare and review the gold-standard data, and their time costs money too. Add inference costs: a fine-tuned open model has to run somewhere. Ask the contractor to break the estimate down into these four items, rather than quoting a single total.
How long does fine-tuning take?
The training run itself takes hours; the full project takes weeks. Yandex AI Studio estimates fine-tuning at 1 hour to 1 day, depending on data volume and system load. The longest steps are collecting and approving gold-standard examples, several rounds of “train — check — fix the data,” and validation on a holdout set. Plan timelines based on when the dataset is ready, not when the GPU starts.
What hardware do you need to fine-tune a model in-house?
For LoRA and QLoRA, a single server GPU is enough for mid-sized models; full fine-tuning of large models requires a cluster. The authors of QLoRA (Dettmers et al., 2023) showed fine-tuning a 65-billion-parameter model on a single GPU with 48 GB of memory thanks to 4-bit quantization. The real requirement depends on model size, example length, and the method you choose. More on hardware costs in the 2026 local LLM review.
Questions about platforms and models
Can you fine-tune ChatGPT through the OpenAI API?
Not for new customers anymore. OpenAI is phasing out independent fine-tuning: as of May 7, 2026, it is closed to organizations that had not started it before; as of July 2, 2026, it is closed for those who had not used fine-tuned models; and on January 6, 2027, new jobs will also be closed for existing customers. Already fine-tuned models will continue to work until their base models are retired.
Can you fine-tune YandexGPT or GigaChat?
Yes, both have such features, but the terms need to be checked on the project date. Yandex AI Studio describes fine-tuning with LoRA for YandexGPT Lite, classifiers based on it, and an embedding model, with a dataset up to 5 GB; on October 19, 2026, the page stated that the feature was disabled due to an incident. The GigaChat documentation includes a fine-tuning section; confirm the current dataset requirements with Sber.
Which open models can be fine-tuned for commercial use?
Many can, if you comply with the license for the specific model. For example, Qwen3-8B is published under Apache 2.0 — a permissive license. The Llama 3.1 license allows commercial use, but requires separate permission from Meta if the company has more than 700 million monthly active users, and it sets conditions for naming derivative models. Keep a copy of the license together with the model version in the project documentation.
Questions about quality and operations
How can you tell whether fine-tuning helped?
Compare the base model and the fine-tuned model on the same holdout set of real requests using zaranee agreed criteria. For classification and field extraction, count the share of correct answers; for text outputs, use expert evaluation against a checklist. Also check for regressions: did the model get worse at general tasks? We described how to build such a set in the article “Your own benchmark for choosing an LLM”.
What happens to a fine-tuned model when a new version of the base model is released?
The adapter is tied to a specific version of the base model, so when it is replaced, fine-tuning is usually repeated. With cloud providers, a fine-tuned model lives as long as the base model does: OpenAI explicitly ties the shutdown of fine-tuned models to the retirement of their base. Keep the dataset, scripts, and validation set — then retraining will take days, not months. More on version transitions in the article about LLM model deprecations.
Can fine-tuning make a model worse?
Yes. A narrow or repetitive dataset can make the model worse on tasks it never saw during training, and it will confidently reproduce mistakes in the gold data. Fine-tuning on answers with made-up facts increases the risk of hallucinations. Protection comes from clean, consistent examples, a holdout set, testing on general tasks, and the ability to quickly roll back to the base model. More on data preparation in the article about data for AI implementation.
What fine-tuning does and does not do: summary table
| Task | Fine-tuning | What is usually better |
|---|---|---|
| Strict response format (JSON, email template) | Suitable | Try a prompt with a schema first |
| Brand voice and terminology | Suitable | — |
| Classifying inquiries into your own categories | Suitable | Fine-tuning or a classifier |
| Answers based on current documents and pricing | Not suitable | RAG with document search |
| Facts that change frequently | Not suitable | RAG or a query to the source system |
| Working without external cloud services | Suitable together with an open model | Local deployment |
*The table is the editorial team’s methodology, summarizing the sources below, not the result of our own experiment.*
Primary sources and verification method
We used only open primary sources, checked on October 19, 2026:
- OpenAI: Supervised fine-tuning — minimum 10 and start with 50 examples, JSONL format.
- OpenAI: Deprecations — timeline for phasing out independent fine-tuning.
- Yandex AI Studio: Fine-tuning — models, LoRA, timelines, dataset size, feature status.
- Hu et al., LoRA, 2021 and Dettmers et al., QLoRA, 2023 — method efficiency.
- Ovadia et al., Fine-Tuning or Retrieval?, 2023 — comparison of fine-tuning and RAG.
- Qwen3-8B on Hugging Face and the Llama 3.1 license — license terms.
- Federal Law No. 152, Article 6, Article 12, Article 18 and the Civil Code of the Russian Federation, Article 1296.
The figures are based on data from the authors and vendors; we did not independently verify them. Limitations: we could not verify GigaChat’s terms from the primary source, and cloud platform statuses change faster than article updates are published.
How AIrassvet helps with LLM fine-tuning
Fine-tuning pays off when, before the GPU starts, it is clear which model error it will fix and how that will be measured. On the AIrassvet website the following areas are listed to address this task:
- LLM implementation and fine-tuning, including model deployment on the customer’s servers if examples cannot be sent to an external cloud;
- RAG systems built on your company’s knowledge base — for cases where the model is missing facts, not format;
- process audit and data preparation: collecting gold-standard examples, holdout sampling, and testing before and after.
The first step is to choose one process where the model makes mistakes, document the current error rate, data sources, processing constraints, and the acceptance criteria.
Bottom line: measure the error first, then fine-tune
Fine-tuning an LLM on company data is a practical tool for format, style, and classification, but not a way to “load” a knowledge base into the model. To get started, you only need dozens of high-quality examples, and the main risks are in the data: personal information, conflicting ground truth, and the lack of a validation set.
That’s why you should start with a set of real queries and an error baseline, verify the selected platform against its current status, and resolve data-set and adapter rights in advance. Then fine-tuning becomes a controlled experiment rather than a random purchase of GPU hours.
