The Factory

How a model is made

One Line, A Lane Of Your Own

Eight stages, with a Build Log running beneath every one of them, and a separate lane for each customer from the sixth stage onwards.

The first four stages are built once. The fifth is built once for each industry. Only the last three are done for each customer.

Every stage writes to the Build Log. At the end of the line it ships with the model as its assurance pack.

Stage by stage

Eight Stages, One Build Log

1. Sourcing

   Gathering text and documents the business has the right to use, and recording where each item came from and on what terms.

   Technology Licensed and open data sources, secure storage, and the Build Log.

2. Refining

   Cleaning the raw data by removing duplicates, low-quality text and personal data, then converting it into the form a model learns from.

   Technology Data processing pipelines, and automated quality and safety filters.

3. Base Model

   Producing the general-purpose foundation that everything else is built on.

   Technology A cluster of graphics processing units ('GPUs'), the specialist chips models are trained on.

4. Platform Build

   Turning the base model into a family of two or three model sizes, then teaching each to follow instructions, to say when it does not know, and to refuse when it should.

   Technology Distillation, where a larger model teaches a smaller one, and post-training, meaning further training after the base is finished, to shape behaviour.

5. Sector Pack

   Teaching the platform the specific tasks an industry needs done, each with a measurable target.

   Technology Adapters, small add-on layers trained on top of the model without changing the model underneath, plus a scored test set for each task.

6. Customer Add-On

   Teaching the model one organisation's private data, kept wholly apart from every other customer's. The line splits into separate lanes here.

   Technology A separate adapter for each customer, trained in isolation with the data encrypted.

7. Proving

   Testing the finished model against the sector pack's targets and the customer's real tasks, deliberately trying to make it fail, and recording how it behaves when it is unsure or should refuse.

   Technology An evaluation harness: automated test suites that score the model's answers.

8. Shipping

   Shrinking the model so it runs well on the customer's own hardware, including sites with no internet connection, and delivering it with its assurance pack.

   Technology Compression, which makes the model smaller and faster with little loss of quality; a digitally signed package, so the customer can prove it has not been tampered with; and software that runs the model on their own servers.

• The Build Log is started at the first stage and added to at every one after it: what data went in, the right to use it, and what was done to it. At the last stage it is issued as the assurance pack, alongside the model's test results.
• The return loop. As your data changes, your add-on is retrained and re-proven, under your control.

Why a line

Built Once, Made For You

Building a language model from nothing for one organisation is slow and costly. A line lets most of that work be done once.

• The foundation is shared. The first four stages are built once, and every model on the line is made from them.
• Your industry's work is done once. A sector pack teaches the platform the tasks a whole industry needs, each with a measurable target. It holds nothing private, so it can serve a whole industry; what sets your model apart comes next.
• Your model is yours alone. We train your add-on on your data, in its own lane, and nothing in it reaches another customer.
• So the cost is shared as well. Most of what goes into your model is paid for once, across every model on the line.

What ships with every model

A Model, And Its Build Log

Every model leaves the line with an assurance pack: its Build Log, recording what went into the model at every stage, the right to use it and what was done to it, together with how the model performed when it was tested on your real tasks.