AI Answer Library
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.
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.
| Method | Data volume | What it is good for | Main risk |
|---|---|---|---|
| Prompting with few-shot examples (not strictly fine-tuning) | Five to fifty examples | Testing whether the need is real, on the same day | A swollen prompt is paid for on every call and drifts as it grows |
| Parameter-efficient fine-tuning (LoRA and similar) | Hundreds to a few thousand | Locking output structure, unifying terminology and tone | Too few or too homogeneous samples overfit; it barely helps with new knowledge |
| Full-parameter supervised fine-tuning | Tens of thousands and up | Shifting wholesale into a markedly different domain register | Expensive in compute and data; can degrade general ability, so regression-test it |
| Continued pre-training on domain corpora | From hundreds of millions of tokens | Domains whose corpus diverges sharply from general text, such as specialised industrial or scientific literature | Costed at cluster scale; for most companies the return does not justify it |
| Preference alignment (DPO-style) | A few thousand preference pairs | Choosing between two answers that are both correct but not equally good | Worthless without consistent annotation standards — agreement matters more than volume |
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.
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
People also ask
Should an enterprise knowledge base use RAG or a fine-tuned LLM?
How accurate can an enterprise RAG knowledge base actually be?
What server specs do you need to self-host a large language model?
Digital Doctor
Turn a physician’s expertise and clinical experience into a digital asset they own 100%, with an AI Digital Doctor handling health consultation and medical education 24/7.
Siemens SCL Auto-Generation Agent
Turn "requirements + process notes + I/O table" into Siemens SCL code automatically — a 10× lever for electrical engineers.