[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"guide-what-is-fine-tuning-and-when-you-actually-need-it::en":3,"guide-related-what-is-fine-tuning-and-when-you-actually-need-it::en":18},{"slug":4,"title":5,"excerpt":6,"body":7,"meta_title":8,"meta_description":9,"keywords":10,"category":16,"published_at":17,"updated_at":17},"what-is-fine-tuning-and-when-you-actually-need-it","What Is Fine-Tuning? And When You Actually Need It","Fine-tuning adapts a model's behaviour by training it on your examples. This guide explains what it does well, what it cannot do, and why retrieval or a better prompt solves most of the problems people reach for it to fix.","\u003Ch2>What fine-tuning actually does\u003C\u002Fh2>\n\u003Cp>Fine-tuning takes a model that has already been trained on a very large corpus and continues training it on a much smaller set of your own examples. Each example is a pair: an input, and the output you want. After enough of them, the model's default behaviour shifts toward the pattern in your data. Nothing about the process is mysterious, and nothing about it is magic — you are nudging a set of weights that were already shaped by orders of magnitude more text than you will ever supply.\u003C\u002Fp>\n\u003Ch2>The distinction that saves the most money\u003C\u002Fh2>\n\u003Cp>Fine-tuning is very good at teaching a model how to respond and comparatively poor at teaching it what is true. Format, tone, structure, a consistent taxonomy, a house style, a narrow classification decision — these are behaviours, and a few hundred well-chosen examples can move them decisively. Facts about your product, your customers or last week's policy change are a different problem. They change, they are numerous, and a model that has absorbed them into its weights cannot tell you where an answer came from or be corrected without another training run.\u003C\u002Fp>\n\u003Cp>That is why the standard answer for knowledge is retrieval: keep the facts in a system you can update, fetch the relevant ones at request time, and put them in the prompt. The two techniques are not rivals. A production feature often fine-tunes for the shape of the answer and retrieves for its content.\u003C\u002Fp>\n\u003Ch2>Try the cheaper things first, in order\u003C\u002Fh2>\n\u003Cp>Most teams that believe they need fine-tuning have not exhausted the alternatives. Work through them in cost order. First, a clearer prompt: an explicit description of the task, the audience and the output format fixes more problems than anything else, and costs an afternoon. Second, examples in the prompt — showing three or four ideal outputs is the fastest way to communicate a format. Third, retrieval, if the failures are about missing knowledge. Fourth, decomposition: splitting one prompt that does three things into three calls that each do one usually beats any amount of tuning.\u003C\u002Fp>\n\u003Cp>Only after those still fall short is fine-tuning the right lever — typically when you need the behaviour to be consistent across thousands of calls, when the instruction needed to get it via prompting has grown long enough to cost real money per request, or when you want a smaller, cheaper model to do a narrow job that currently requires a larger one.\u003C\u002Fp>\n\u003Ch2>What it costs beyond the training bill\u003C\u002Fh2>\n\u003Cp>The compute is often the smallest line. The dataset is the real work: examples have to be collected, cleaned, made consistent, and held to a standard, because a model trained on contradictory examples learns to be inconsistent. You need a held-out set you never train on, or you will have no way to tell whether the tuned model is better or merely different. And a fine-tuned model is a new artefact to own — it has a version, it needs re-evaluation whenever the base model is updated, and it can quietly become the reason you cannot move to a newer, cheaper model later.\u003C\u002Fp>\n\u003Ch2>How to decide, concretely\u003C\u002Fh2>\n\u003Cp>Write down the failure you want to fix, in one sentence, with three real examples of it. Then ask which category it falls into. If the model does not know something, that is retrieval. If the model knows but answers in the wrong shape, try prompting and examples first, then fine-tuning. If the model is right but too slow or too expensive, fine-tuning a smaller model is a legitimate answer. If the failures have no pattern you can describe, no amount of tuning will help, because you cannot build a dataset for a target you cannot state.\u003C\u002Fp>\n\u003Ch2>If you go ahead\u003C\u002Fh2>\n\u003Cp>Start with an evaluation set before you start with training data — a few dozen realistic inputs and the outputs you would accept, scored the same way every time. Without it you will be comparing impressions. Keep the training set small at first and see whether the curve is moving; a large dataset assembled before any signal is a long detour. Record which base model, which data version and which settings produced each tuned model, because you will need to reproduce it. And re-run your evaluation against the plain base model periodically: the honest outcome of many fine-tuning projects is that a newer general model has caught up, and the maintained artefact is no longer earning its keep.\u003C\u002Fp>","What Is Fine-Tuning?","What fine-tuning is, what it changes and what it does not: why it teaches format and style rather than facts, and when retrieval or prompting is the better answer.",[11,12,13,14,15],"fine-tuning","llm customization","rag vs fine-tuning","model training","ai features","fundamentals","2026-08-10T03:45:02+00:00",[19,24,28,32,37,42],{"slug":20,"title":21,"excerpt":22,"updated_at":23},"ai-tool-pricing-models-seat-vs-usage-vs-credits","AI Tool Pricing Models: Seat-Based vs Usage-Based vs Credits","The three common ways AI tools charge — per seat, per usage, and by credits — and how to reason about which one will actually be cheaper for the way your team works.","2026-08-05T14:32:26+00:00",{"slug":25,"title":26,"excerpt":27,"updated_at":23},"how-ai-image-generators-differ-diffusion-vs-the-rest","How AI Image Generators Differ: Diffusion vs the Rest, in Plain Terms","A non-technical explanation of how AI image generators work, why the diffusion approach became dominant, and what practical differences to expect between tools.",{"slug":29,"title":30,"excerpt":31,"updated_at":23},"how-to-automate-your-workflow-without-code","How to Automate Your Workflow Without Code","A practical sequence for building automations that survive: picking the right process, mapping it before touching a tool, and handling the failure cases that break most first attempts.",{"slug":33,"title":34,"excerpt":35,"updated_at":36},"how-to-build-a-chatbot-without-coding","How to Build a Chatbot Without Coding","A practical route to a working chatbot using no-code tools: deciding scope, connecting your own content, handling the questions it cannot answer, and knowing what it will cost.","2026-08-05T14:32:27+00:00",{"slug":38,"title":39,"excerpt":40,"updated_at":41},"how-to-change-a-prompt-without-breaking-production","How to Change a Prompt Without Breaking Production","Prompts get edited in a text box and shipped in seconds, which is why they break things quietly: no compiler, no stack trace, no obvious moment of failure. Give them the release discipline code gets.","2026-08-24T03:30:02+00:00",{"slug":43,"title":44,"excerpt":45,"updated_at":23},"how-to-choose-an-ai-writing-assistant","How to Choose an AI Writing Assistant","A practical framework for picking an AI writing tool — matching it to the kind of writing you actually do, checking editing controls, and avoiding tools that produce confident but generic copy."]