Menu

Use case

Which conversations should I fine-tune on?

Fine-tuning amplifies whatever you feed it, including mistakes. Measure which examples actually improve the agent before they become weights.

The problem

Fine-tuning runs cost money and bake data in. Teams pick training examples by feel, every bad example becomes a learned behavior that is expensive to unlearn.

The workflow

  1. Baseline

    Score the agent over your eval set before any tuning.

  2. Value

    Measure each candidate example's marginal effect with a confidence interval, against fresh controls.

  3. Select

    Build the training portfolio under budget; regressions and redundant examples are rejected with a reason.

  4. Export

    Render the selected examples as a tuning-set JSONL, re-exporting is byte-identical, and export never mutates a destination.

What it looks like

kno
Export run 20260828T233415-8f3a1b2c4d5e (completed)
  destination  tuning_set
  wrote        tuning.jsonl (2 assets, 1024 bytes)
  manifest     tuning.jsonl.manifest.md

The recipe

Turn production transcripts into an eval set with kno mine, then value candidate training examples against it the same way as any other asset.

What you get

A tuning-set JSONL of measured keepers with a manifest. Note: today Kno exports the set and your existing training pipeline consumes it; the Tuner and the proxy-FT bridge arrive in v0.2.

Stages involved

Try it →