Fine-Tuning
Advanced 13 min read Fine-Tuning

Eval After Tuning

The verdict on tuned behavior

Post-tuning eval answers two questions: did it learn the target AND did anything else regress. A dedicated suite — behavioral, cross-domain, adversarial — is the only honest verdict on a tuned model.

post-tuning evalregressioncross-domainbehavioraladversarialholdout
Prerequisites: Tuning Pipeline
Before we start

What you will learn

  • Grade the tuned behavior precisely
  • Probe unrelated skills for regressions
  • Compare against the untuned baseline
Introduction

What is this?

Post-tuning evaluation grades the target behavior on held-out data (never the training set), checks unrelated capabilities for regressions, runs adversarial/format tests, and compares the whole picture against the untuned baseline. That comparison is the decision: promote, retune, or revert.

Motivation

Why do you need it?

  • Training loss lies; held-out tells.
  • Tuning can silently damage other skills.
  • Adversarial cases catch format/instruction regressions.
  • Baseline comparison makes the verdict a decision, not a feeling.
Think about it

Mental model

Imagine

Final exams + an upkeep audit

Think of

Post-tuning eval

A student who aces the new subject but flunks every previous course (regression) isn't a success. Exams assess the new material; a transcript audit checks the old; and both compare to last term's baseline. Only the full picture decides the grade.

Mechanics

How it works

  • Freeze the held-out behavioral set.
  • Run target-behavior grading (rubric / exact).
  • Run the regression suite on unrelated skills.
  • Add adversarial format/tool cases.
  • Tabulate vs untuned baseline across all dimensions.

Verdict matrix.

Behavior
Regression
Adversarial
Baseline
Decide
Real world

In production

  • Support-tone tune judged on held-out tickets + general QA.
  • Tool-intent tunes checked against the tool-agnostic eval.
  • Every promotion posts the matrix to the review PR.
Guidelines

Best practices

  • Hold-out data stays out of training by construction.
  • Keep a cross-domain suite for every tuned model.
  • Publish the matrix (target vs regression) with each promotion.
  • Re-run the same suite after base-model upgrades.
Watch out

Common mistakes

  • Evaling on training samples (memorization applause).
  • Only testing the target behavior.
  • No baseline row — 'passing' means nothing in context.
Root cause analysis

Common issues

Error
Eval row echoes training
Root cause

Dataset leak

Fix

Rebuild splits with airtight holdout.

Error
Regressions hidden in noise
Root cause

Unrelated suite too small

Fix

Size the cross-domain suite to catch 5-point shifts.

When things break

Debugging tips

  • Target great, regressions → adapters bled; restrict targets.
  • Target weak → data quality or capacity; strengthen dataset.
  • Baseline beats tuned → the lesson was promptable; don't ship.
Interviews

Interview checkpoint

Test yourself

Quick quiz

0/2
1

Training-set performance tells you…

2

Post-tuning eval should include unrelated skills.

Go deeper

Official documentation

Fine-tuning evaluation (OpenAI)
docs