Back to Docs

Fine‑tuning primer

When to fine‑tune, what you need, and how Meridian makes it fast.

Why fine‑tune?

Prompt engineering gets you 80% of the way. Fine‑tuning closes the last 20% — tone, format adherence, domain vocabulary, and edge‑case reliability. If you find yourself repeating the same instructions in every prompt, it is time to bake them into the model weights.

What you need

  • 50–100 high‑quality examples covering happy paths and common failure modes.
  • A consistent schema — every example must follow the same system → user → assistant structure.
  • A held‑out eval set (10–20 examples) to measure improvement objectively.

Meridian workflow

  1. Upload dataset — drag a JSONL file into the Meridian dashboard. We validate schema, deduplicate, and flag low‑quality rows.
  2. Configure run — choose base model, epoch count, and learning rate. Sensible defaults are pre‑selected.
  3. Train — Meridian provisions GPU capacity, runs the job, and streams progress. Typical turnaround is under 90 minutes.
  4. Evaluate — compare base vs. fine‑tuned side‑by‑side on your eval set. Ship when the delta justifies the cost.

Ready to start? Head to the step‑by‑step walkthrough for a concrete example with real data.