A few thousand examples are written by generating them, and generated examples carry the generator's bias. The way to keep them honest is to generate from varied seeds, filter with a check you can read, and keep the ones that survive.
About 1 min of reading
Measure the filter.
Survival
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
Generating examples with the model you are about to tune teaches it to copy itself. Use a different, stronger model for generation, or write them, and keep that separation deliberate.
How many generated examples passed the filter. A set where ninety percent failed is a generator that is not producing what you need, and the rate is the number that says so.
Generate from real inputs, from paraphrases, and from the rare cases you collected by hand. Generating from one template produces a set that is one shape wide, which is the most common failure of a synthetic set.
Reject outputs that fail a check: wrong label set, invalid JSON, too short, contradicts the input. A filter you can read is what keeps a synthetic set from being a wall of plausible noise.
A few hundred hand written examples set the quality bar; the generated ones fill the range. A set made only of generated examples has no bar, and its quality is whatever the generator happened to do.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.