All servicesService 03 / 03Fine-tuning / Evaluation / Adapters

Model tuning

I adapt a language model to your categories, terminology or output format, then compare it with the original on examples it has never seen.

Typical length
3 to 6 weeks
Format
Build engagement
See what is included
A field of pins arranged at different heights, shown as a dot mesh.
(01) Overview

Overview


A general model knows a lot about the world and very little about your categories. Prompting closes part of that gap. Tuning on examples labelled by your team can close more of it, but only a fair comparison tells you whether it was worth the effort.

I start with a baseline and a held-out test set, so every change is measured against the same examples. You receive the tuned model, the numbers and the mistakes that remain, in plain language.

(02) Example use

A model that knows your ticket categories


Two support tickets mention a payment. One is a billing dispute and the other is a checkout bug. I use examples labelled by your team to teach the distinction, then measure how often the model assigns the right label on a separate set of tickets.

  1. 01Labelled examples
  2. 02Baseline prompt
  3. 03Tuned adapter
  4. 04Held-out test
  5. 05Comparison report
(03) Deliverables

What you receive.


D/01

A baseline and test set

We agree on what a good answer looks like, measure the current model with a prompt, and reserve examples for the final comparison.

D/02

A tuned model or adapter

I prepare your training examples and adapt a suitable base model. A small adapter can carry the learned changes while keeping the original model weights fixed.

D/03

An evaluation and handover

You receive the comparison on unseen examples, the mistakes that still need attention and instructions for running the tuned version.

Further reading: Hugging Face on parameter-efficient fine-tuning
(04) Process

Four steps, one task.


  1. 01

    Define

    Agree on labels, a scoring rule and a held-out test set.

  2. 02

    Baseline

    Measure the best prompt on an unchanged model first.

  3. 03

    Tune

    Prepare examples and train an adapter on a suitable base model.

  4. 04

    Compare

    Report results on unseen examples and the errors that remain.

Where to startBring sample inputs, the answers you want and a few outputs that miss the mark today. Those examples help establish whether tuning is worth the training effort.

(05) Lab

Train a tiny word model.


neu-lab / model-tuningRuns locally

Train a tiny model.

This small model learns by counting word pairs. Edit its training sentences to see how examples change the words it predicts after “please”, “we” or “your”.

Base model

  • check50%
  • open25%
  • share25%

Train the model to compare distributions.

(06) Questions

Common questions.


It depends on the task. A few hundred clean examples often show whether tuning helps. The baseline tells us early.

No. Sometimes a better prompt wins, and the report will say so. That is why the baseline comes first.

On a provider that supports tuning, or on your own infrastructure with an open-weight model.

Start a brief for model tuningBring one taskNext / 01Agentic systems