AI Answer Library

How much data do you need to fine-tune a domain-specific LLM?

Short answer

It depends on the method, and the answers span three orders of magnitude. A lightweight fine-tune (LoRA-style) for tone and output format needs hundreds to a few thousand curated samples. Getting a model to reliably internalise a whole domain style usually takes tens of thousands. Continued pre-training to inject domain knowledge is measured in hundreds of millions of tokens, and most companies should not attempt it. All of them share one precondition: the samples must be human-verified — dumping raw historical tickets in just teaches the model your past mistakes.

Key points

  • 01The question is usually asked in the wrong order. First decide whether the problem is format, style or knowledge; the data volume follows from the method, not the other way round.
  • 02Quality outruns quantity by a wide margin. A few hundred human-checked samples covering typical and edge cases usually beat tens of thousands of uncleaned historical records.
  • 03Fine-tuning does not reliably teach facts. If the model needs current policies, prices or inventory, that is RAG territory — forcing facts into weights is the most expensive and least reliable route.
  • 04Distribution beats raw count. A dataset made only of easy, common questions improves nothing on hard cases and can make the model more confidently wrong.
  • 05Budget the data before the compute. Preparing it usually takes far longer than the training run, and it is the one part no hardware vendor can do for you.

Data volume by fine-tuning method

The ranges below come from common open-source fine-tuning practice, framed around Chinese business text and a single task type. Use them for a first pass, not for precise planning: the real requirement scales with task difficulty, base-model strength and sample diversity. The same thousand examples can be more than enough to standardise ticket-classification output and nowhere near enough to teach a model to draft compliant legal opinions.

MethodData volumeWhat it is good forMain risk
Prompting with few-shot examples (not strictly fine-tuning)Five to fifty examplesTesting whether the need is real, on the same dayA swollen prompt is paid for on every call and drifts as it grows
Parameter-efficient fine-tuning (LoRA and similar)Hundreds to a few thousandLocking output structure, unifying terminology and toneToo few or too homogeneous samples overfit; it barely helps with new knowledge
Full-parameter supervised fine-tuningTens of thousands and upShifting wholesale into a markedly different domain registerExpensive in compute and data; can degrade general ability, so regression-test it
Continued pre-training on domain corporaFrom hundreds of millions of tokensDomains whose corpus diverges sharply from general text, such as specialised industrial or scientific literatureCosted at cluster scale; for most companies the return does not justify it
Preference alignment (DPO-style)A few thousand preference pairsChoosing between two answers that are both correct but not equally goodWorthless without consistent annotation standards — agreement matters more than volume

Where the data comes from, and what makes it usable

A first fine-tuning dataset typically comes from three places: historical tickets and support conversations, internally reviewed standard documents, and samples annotated on the spot by domain experts. The first two are plentiful but dirty; the third is clean but slow. A workable approach mixes them: filter historical records down to cases judged as correctly handled, then have experts rewrite each one into a reference answer. Three tests decide whether a sample is usable — the input is shaped like real traffic rather than an idealised question; the output is exactly how you want the model to answer every time, not merely acceptable; and comparable questions are answered consistently across the whole dataset. The third is the easiest to overlook and the most damaging: three different answering styles for one question type teaches the model to pick one at random.

Four checks before you start fine-tuning

One: confirm fine-tuning is actually the right tool. Most complaints of the form "the model does not understand our business" decompose into retrieval failures that RAG fixes the same day and fine-tuning cannot. Two: have an independent evaluation set, strictly disjoint from the training data and built by the business side rather than by whoever trains the model — otherwise all you can prove is memorisation. Three: run regression tests. Domain fine-tuning squeezes general capability, so re-check general tasks afterwards for visible degradation. Four: keep a rollback. Ship the fine-tuned model alongside the previous version so you can switch back immediately if results disappoint. With all four in place, even a mediocre first attempt tells you what to adjust next; missing any one of them, you cannot attribute the outcome either way.

Where this applies

When this answer does not hold

  • These ranges are empirical values from common open-source practice, framed around Chinese business text and a single task type. Task difficulty, base-model strength and sample diversity all shift the real requirement substantially.
  • If the underlying need is knowledge that changes — policies, prices, inventory — this answer does not apply. That belongs to RAG, and no amount of fine-tuning data guarantees freshness.
  • If the training data contains personal information, customer records or protected third-party content, complete redaction and authorisation first. Compliance obligations do not vanish because the data is only used internally.
  • This assumes self-hostable open model families. If you can only fine-tune through a closed API, the available methods, data formats and ceilings are all bounded by what the vendor exposes.

People also ask

  • How many samples are enough to fine-tune an LLM?
  • How many samples does a LoRA fine-tune need?
  • Can a few hundred examples produce a usable fine-tune?
  • Where does the training data for a domain model come from?
  • For fine-tuning, does data quality or quantity matter more?
Written by: YGG Technology Solutions TeamPublished: 2026-08-01Last reviewed: 2026-08-01