InstructLab: Teaching a Model What Only Your Shop Knows

Every AS/400 shop we have ever looked at has the same quiet problem, and it has nothing to do with technology. There is one person who knows why the month-end job has to run in that exact order. There is no document. There never was. And that person is a few years from retiring.

InstructLab is the most interesting attempt yet at making it worth writing that down. It is an open source project from IBM Research and Red Hat, launched in May 2024, and its whole design goal is to let someone who knows the subject teach a model, without that person needing to know anything about machine learning.

The problem it is solving

A general model knows a great deal about the world and nothing whatsoever about your world. It has never seen your part numbering scheme. It does not know that customer type 7 means something different since the 2019 merger. Asking it about your operations gets you a confident, fluent, wrong answer.

Teaching a model new material used to require thousands of carefully prepared examples and a team who knew what they were doing. That cost is why almost nobody did it. InstructLab's method, which came out of IBM Research and is called LAB, short for Large-scale Alignment for chatBots, attacks that cost directly.

How it works, in three steps

1. Somebody writes down what they know

   A handful of examples, not thousands. They go into a structured folder arrangement InstructLab calls a taxonomy, split between knowledge, meaning facts about your world, and skills, meaning things you want the model to be able to do.

2. The tool multiplies those examples

   InstructLab generates a large set of synthetic training examples from the small human-written set. This is the expensive step that used to need a research team, and it is the reason the project exists.

3. The model gets tuned

   The result is your own version of an open model that knows your material. It runs on your hardware. Nothing was sent anywhere, and nobody else got a copy of what your retiring expert knows.

Red Hat packages all of this inside Red Hat Enterprise Linux AI, where it arrives already fitted to an IBM Granite model. That removes the assembly work, which for a shop with no existing Linux AI experience is most of the difficulty.

Why this sits on a hardware site

Because tuning is the single AI task where sending your material to somebody else's cloud is hardest to defend. Model tuning consumes exactly the thing you would least like to hand over: the accumulated operational knowledge that makes your business work the way it does.

So the sequence lines up neatly with what an AS/400 shop already owns. An open Granite model under Apache 2.0, InstructLab to teach it, and a Power server you already paid for. The knowledge, the model, and the machine all stay on your floor. No part of that arrangement involves a monthly meter or a renewal negotiation.

It also means the tuning work you do on the processor you already own carries forward. If the pilot proves out and you later buy a Spyre Accelerator, the model you tuned goes with you. The cargo does not change when the vehicle does.

The honest limits

There is a language point worth knowing too. InstructLab supports Spanish, German, French, and Italian, with Japanese, Hindi, and Korean noted as planned. For a shop with operations in more than one country, that is the difference between one tuned model and several.

Where to start if this interests you

Start with the smallest genuinely useful thing. Not the month-end job. Pick one procedure that three people ask about every week, write fifteen honest examples of the questions and the right answers, and tune a small Granite model on your existing Power hardware using the on-chip Matrix Math Acceleration you already have.

If that works, you have proved the method on something small and you have the beginnings of a habit of writing knowledge down. If it does not work, you have spent a fortnight and no capital.