Fine-Tuning
Advanced 12 min read Fine-Tuning

Managing Model Updates

Surfing the base-model cadence

Base-model upgrades change your tuned adapters' world: re-evaluate, re-tune, or re-baseline deliberately. Update management is a planned cadence with rollback, not a fire drill.

model updatebase upgradere-tunerollbackversioningenrich
Prerequisites: Eval After Tuning
Before we start

What you will learn

  • Evaluate upgrades before adoption
  • Re-tune adapters on new bases
  • Version and roll back cleanly
Introduction

What is this?

Model updates arrive on a cadence you don't control. Managed adoption: evaluate the new base against the same suites (zero-shot + your tuned adapters), decide re-tune vs re-baseline, pin the whole stack (base + adapter + config), and keep an instant rollback path. Unmanaged upgrades silently move your evals.

Motivation

Why do you need it?

  • New bases change behavior — even without tuning.
  • Adapters trained on old bases may degrade or improve unpredictably.
  • Versioned stacks make compare and rollback trivial.
  • Scheduled evaluation prevents surprise regressions in prod.
Think about it

Mental model

Imagine

Rebalancing your portfolio on a schedule

Think of

Model update management

You don't panic-sell when indexes shift; a quarterly rebalance re-tests the mix (re-eval), adjusts exposure (re-tune/baseline), and keeps records so you can revert a bad move (rollback). Timid or eager, the schedule keeps decisions deliberate.

Mechanics

How it works

  • Maintain a pinned 'champion stack' (base + adapter + config).
  • On a new base release, run the full eval matrix.
  • Compare zero-shot + existing adapters vs champion.
  • Re-tune adapters if behavior shifts matter.
  • Promote or keep champion; rollback switch is instant.

Update cadence.

Release
Eval
Decide
Pin
Rollback
Real world

In production

  • Quarterly base-upgrade review gate.
  • Adapters re-trained on the new base when adopted.
  • Stack registry versioned so any prod snapshot is restorable.
Guidelines

Best practices

  • Never auto-upgrade a base in prod (unreviewed).
  • Re-run the same eval matrix on every release.
  • Keep old champion stacks deployable for rollback.
  • Document what changed behaviorally per upgrade.
Watch out

Common mistakes

  • Grabbing the newest model hot (drift without eval).
  • Re-tuning against a moving base by feel.
  • No pinned stack — can't roll back to 'what worked'.
Root cause analysis

Common issues

Error
New base regressed a tuned skill
Root cause

Adapter not re-validated

Fix

Re-run adapter eval; re-tune or pin old base.

Error
Rollback impossible
Root cause

No versioned stack

Fix

Start pinning now; restore from registry.

When things break

Debugging tips

  • Prod scores shift after release → an unmanaged base upgrade happened; audit.
  • Adapter quality drops on new base → re-tune timing; roll back or redo.
  • Stacks confusing → registry tagging (base+adapter+config hash).
Interviews

Interview checkpoint

Practice

Hands-on lab

A/B Script: Adapter vs Untuned Baseline
30 min · advanced
In progress
Objectives
  • Run both variants over a golden set
  • Score per-dimension deltas
  • Emit a promote verdict

A/B Script: Adapter vs Untuned Baseline

30 min · advanced

0/4 steps
0%
sandbox · Linux sandbox
$
Try these commands
Test yourself

Quick quiz

0/2
1

New base models should be adopted…

2

Keeping old stacks deployable aids rollback.

Go deeper

Official documentation

Model versions (OpenAI)
docs