9/24/2025
Published on Convert on 24/09/2025.
In Part 1, the 0.6B model only gave two insights; the 1.7B model was close to ChatGPT. This post gets the 0.6B model to “perfect” by changing how we prompt.
- What didn’t fix it: Tweaking temperature and adding more few-shot examples helped a bit but not enough.
- What worked: breakpoints. Like pausing in a debugger, we split the job so the model isn’t asked to do everything at once:
- Prompt 1: List negative insights only (and ignore things like “pigs might fly”).
- Prompt 2: List positive insights.
- Prompt 3: Turn the lists into structured output (e.g. JSON).
- Result: The tiny model matched ChatGPT’s insight quality for the test—with three prompts instead of one, but very fast on a local machine.
The article has full prompt examples and next steps (automation and training).
