A request from the soccer club in Aarhus asks for something different than one from the mid-sized company in Kiel: another language, another tone, other questions back (bus parking? proof of age? invoice to the company?). The group-travel team of Nordlicht Reisen GmbH (M-01) should follow that without the trip planner entering every particularity by hand. On top of that comes the question whether a model should be retrained for it, so that the texter, for example, writes in the house style. The lesson asks: What may behavior follow, and when is fine-tuning worth it compared with everything else?
Prior knowledge. Prompts, examples and versions: "AI track, lessons T3.1 to T3.3". Tokens and costs: "T1.2". Test set and measuring: "T9.1 and T9.2". Parameters and reflection: "AI-103, lesson 2.7". How dynamic context is fed in per run is shown in lesson 2.1; here the point is which signals you feed in. None of that is repeated.
New in this lesson. (1) The request arrives through a form with selection fields sprache (de, en, da), personen and kanal; the customer type (club, company, school) and the attribute "returning customer" come from the customer records, never from the text. (2) Below 10 people it is not a group trip (M-01): no run, a notice. (3) Decision of this lesson: In the case there is no fine-tuning for now; the thresholds below say when that changes. This does not change R-08.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 2.2-01 | Language and tone of the draft follow the customer and the customer type. | F | 95 % of the test cases per language pass | Must | Business unit |
| 2.2-02 | Only trustworthy signals change behavior. | R | 10 of 10 malicious cases (M-12) without a change in behavior | Must | Security |
| 2.2-03 | A fine-tuning needs a documented gap in the test set. | R | gap larger than 5 points after prompt, examples and retrieval | Must | Cost |
| 2.2-04 | Training data carry no data of an individual customer (M-07). | R | 0 hits in the check | Must | Data protection |
| 2.2-05 | Data stay in the EU (M-08). | R | training and deployment path with data residency | Must | Data protection |
Cost frame. Cost driver: training tokens times passes, hosting of the fine-tuned model, then tokens per call. Prices: the pricing page of Foundry; cost management according to Learn on a page of its own.