Decide honestly whether a business needs fine-tuning, and prove it

The situation

An online clothing store in Lahore wants a model that writes product descriptions in its own style: short, warm, mixing a few Urdu words, always mentioning fabric and care. A vendor has quoted them for fine-tuning. The owner wants to know first whether a good prompt could do the same job for nothing.

Try the cheap way properly, prepare data as if you would fine-tune, and decide on evidence. If you have access to a fine-tuning tool, run it; if not, complete every step up to training and the evaluation plan.

What you'll be able to show

• Rule out fine-tuning first by improving the prompt with examples
• Prepare consistent training data and hold back a test set
• Compare models fairly on the same test set and the same prompt
• Make a cost-and-maintenance decision the owner can follow

What you need

• Any AI assistant for the prompt baseline
• A spreadsheet for the dataset
• Optional: a fine-tuning service (for example OpenAI fine-tuning or a hosted open model) if you have access

Tasks

1. 1Write the target style down

   Write ten product descriptions by hand in the store's style (invented products are fine). Then write the style as five rules a stranger could follow.
   A correct result: Ten descriptions and a five-rule style guide.

2. 2Try the prompt first

   Build a prompt with the five rules and three of your examples. Generate descriptions for ten new products. Score each against the five rules (1 point per rule met).
   A correct result: Ten generated descriptions with scores and an average out of 5.

3. 3Prepare the dataset

   Write or collect at least 50 input-output pairs (product facts in, description out). Check twenty at random for consistency with the style guide and fix any that break it.
   A correct result: A dataset of at least 50 pairs and a note of the inconsistencies you fixed.

4. 4Hold back a test set

   Set aside 20 percent of the pairs as a test set before any training. Never use these in the prompt or in training.
   A correct result: Two files, training and test, with the counts.

5. 5Compare fairly

   Run the test set through the prompt-only approach and, if available, the fine-tuned model, using the same instructions. Score both with the same five-rule rubric.
   A correct result: A comparison table of average scores. If no fine-tuning access, the prompt score and a written evaluation plan instead.

6. 6Decide

   Write a half-page recommendation for the owner: which approach, the cost of each (training fees, per-use cost, time), how often it would need redoing as base models improve, and the date to re-check.
   A correct result: A clear recommendation with costs and a re-check date.

What to hand in

The style guide, the prompt baseline with scores, the dataset split into training and test, the comparison (or evaluation plan), and the recommendation.

How it is marked