Fine-tuning

You show a model several hundred examples of a job done the way you want it done, and it learns the pattern. Afterwards it does that job without being told how each time. It is good at teaching tone, format, and vocabulary. It is unreliable at teaching new facts: if you want the model to know your product catalog, build a search index over your documents instead, which is cheaper and more accurate.

Between 200 and 1,000 for most jobs. Below 200 the model tends to memorize the examples rather than learn the pattern behind them. Above about 2,000 the improvement flattens out for tone-and-format work. Quality matters much more than quantity. Three hundred carefully checked examples beat three thousand messy ones, every time.

Between $0.25 and $2 for most jobs. A LoRA fine-tune of Llama 3.1 8B on a thousand examples takes about 40 minutes on an RTX 4090 at $0.45 per hour, so around $0.30. Larger models and larger datasets cost more but stay in single dollars. The expensive part of fine-tuning is preparing good examples, not the GPU time.

You do. The adapter you produce downloads to your machine and is yours to use, sell, or deploy anywhere. We do not use it or your training data for anything. The one thing to check is the base model license: a model fine-tuned from Llama is still bound by the Llama Community License when you deploy it, whereas one fine-tuned from Mistral 7B under Apache 2.0 carries no conditions at all.

Not for the standard path. Upload a spreadsheet with two columns, one for the input and one for the output you want, set the strength, and collect a finished model. There is a JupyterLab workspace with Axolotl already configured for people who want to watch the training curve and change the settings, but it is an option rather than the route.

That is a normal outcome on a first attempt and it is usually one of three things. Too few examples, so it memorized rather than learned. Inconsistent examples, so it learned to be inconsistent. Or too high a training strength, so it overwrote general ability with your narrow task. The job report shows which is most likely, and a second run with corrected settings costs another dollar or so.