After completing incremental LoRA pretraining for Qwen3-32B with 76,457 documents and roughly 926 million tokens, the author identified a critical problem: when asked about Redis security, the model produced a generic hardening guide instead of the configuration commands and verification steps the user actually needed. This shows that a model’s knowledge is not the same as its ability to answer. The latter depends on the quality and structure of the fine-tuning data.

What this is

Fine-tuning—adjusting a model’s behavior with paired examples so that it imitates a specific response style—addresses not what the model knows, but how it should answer. The path presented here is straightforward: first expand knowledge through pretraining, then use high-quality question-answer pairs to train the model to respond with conclusions, steps, and clear boundaries of applicability.

The author constructed roughly 10,000 cybersecurity question-answer pairs and recommends that teams working in a specialized domain begin with 500–2,000 examples to validate the workflow, then scale to 3,000–10,000 examples covering core scenarios. Each data item should contain at least a question, an answer, and source information: the question should reflect real users and have a single intent; the answer should lead with its conclusion and then break down the details; and the source should be traceable so errors can be traced quickly.

Human involvement remains irreplaceable. Models can expand drafts, but domain professionals must determine which questions are worth asking and which answers look reasonable while leaving important gaps. Letting models generate every question and answer on their own can easily turn existing biases into training standards.

Industry view

We are seeing large-model deployment shift from training bigger models to encoding business experience into data. This is especially relevant for small and medium-sized businesses: rather than spending heavily to retrain general-purpose models, they can compile a manageable set of verifiable question-answer standards around their own products, after-sales support, and compliance workflows.

The risks are equally clear. A dataset of 10,000 high-quality examples is an experimental range, not a universal formula. The more specific an answer becomes, the more likely it is to become obsolete as software versions, contract terms, or security environments change. Without a continuous update mechanism, a model can move from giving irrelevant answers to confidently citing outdated procedures. Human annotation costs, authorization to use data sources, and professional boundaries may therefore become the real bottlenecks.

Impact on regular people

For enterprise IT: The competitive focus of AI projects may shift from model parameters to business-data governance. Organizations need to map high-frequency questions and standard answers first, then decide whether fine-tuning is worthwhile instead of immediately purchasing a larger model.

For individual careers: Prompt-writing skills remain important, but understanding the business, verifying answers, and preserving reusable cases offer greater long-term value. People who can turn tacit experience into clear processes will occupy a more indispensable role in AI projects.

For consumer markets: Vertical AI products will gradually move from broad recommendations to concrete execution, but their reliability depends on whether their data is continuously updated. For decisions involving security, finance, or operations, users will still need to check source material and applicable versions.