Adapting a model on the site's own data
The Adaptation service fine-tunes an open model on the site's own captures and documents, entirely on the appliance. The adapter it produces stays on site and is served behind the same API.
· 4 min read
IN SHORT
Substrate's Adaptation service fine-tunes an open-weight model on data the site already holds, such as camera captures, documents, and records, and does the entire job on the Czar appliance. The output is an adapter that layers on the base model, stays on site, is owned by the customer, and is served behind the same API once the site chooses to load it.
A general model knows what a damaged box looks like in general. It does not know what damage looks like under the lights of your receiving dock, on the packaging your suppliers actually use, photographed by the cameras you actually installed. The gap between the general and the particular is where most of a facility's accuracy problems live.
Substrate's Adaptation service closes that gap by fine-tuning an open model on the site's own data, inside the appliance. This article explains what that means, what it produces, and how the result reaches the floor.
What adaptation is
An open-weight model arrives with general capability. Adaptation adjusts it toward the site's particulars by training on examples the site already has: its own captures, its own documents, its own records of what operators decided. The base model is not rewritten. Instead the service produces an adapter, a compact set of learned changes that layers on the base model at serving time.
Because the adapter is separate from the base, a site can hold several adapters for one base model, serve different ones to different callers, and discard one without touching the others. It can also take a new release of the base model and re-adapt, rather than starting from nothing.
Where it happens
Entirely on the appliance. The data the service trains on is read from where the site keeps it, on the facility network or on the appliance itself. Nothing is uploaded, because there is nowhere to upload it to; Substrate has no default path out of the building, and the Adaptation service does not create one.
This is the practical consequence of running open weights. A model served from someone else's endpoint cannot be adapted on your data without sending them your data. A model whose weights are in your rack can be adapted in your rack.
The data a site adapts on
The useful data is usually already there. Receiving docks have images of every pallet they accepted and every one they rejected. Maintenance teams have manuals, work orders, and the notes technicians wrote after each repair. Operations have transcripts, exception logs, and the decisions supervisors made when a robot handed work back.
Adaptation does not need a labeled dataset assembled from scratch. It needs the site's existing record, organized enough that the service can read it. The integrator's work at this stage is usually curation rather than collection: deciding which captures represent the task and which are noise.
For robot-facing models, the record the pool itself keeps is a source. Every time a vision-language-action model guided a robot and the outcome was recorded, that is an example. A floor that has run for a season has a training set it did not set out to build.
From adapter to the floor
An adapter is not served until the site loads it. Loading places it alongside the base model on the chosen nodes, and from then on callers that ask for it, or that the site routes to it, are answered by the adapted version. The API does not change. A robot vendor's integration that called the base model calls the adapted one without knowing.
The sensible path is to serve both for a while. The site routes a share of requests to the adapted version and the rest to the base, and reads the Evaluation traces for each. Did the adapted perception model refuse fewer unfamiliar items at the pick station. Did the adapted language model answer maintenance questions with better citations. The evidence comes from the site's own floor, which is the only benchmark that matters to it.
When the adapted version is better, the site promotes it. When it is not, the site discards it and the base model keeps serving. Either way the decision was made on the appliance, with the site's data, by the site.
Adaptation as a running practice
The first adapter is rarely the last. Suppliers change their packaging, new items arrive, a line is reconfigured, and the particulars that the adapter learned drift. The Evaluation traces show the drift as rising fallback rates or declining confidence long before anyone on the floor names it.
A site that treats adaptation as a periodic practice rather than a one-time project keeps its models matched to its floor. The ingredients are the same each time: the appliance, the site's recent record, and a comparison in the traces before promotion. Nothing new has to be purchased and nothing has to leave the building.
Who holds the rights to what adaptation produces is the subject of who owns an adapter trained on site. The base models being adapted are described in open-weight models on the appliance. The traces used to compare versions are described in tracing every operation with Evaluation.
Questions this raises
- What kind of data does a site adapt on?
- Whatever it already captures that bears on the task: images of its own items and damage, its own procedures and manuals, transcripts from its own radio channels, and records of decisions its operators made. The data stays where it was, and adaptation reads it on the appliance.
- Does adapting a model take the pool offline?
- Adaptation is a workload on the pool like any other, with a long budget, so it runs in the gaps around serving. A site with heavy serving load may schedule it for quiet hours, and the Grid's routing keeps urgent callers ahead of it.
- How does the site know the adapted model is better?
- By serving both and reading the Evaluation traces. The site can route a share of requests to the adapted version, compare outcomes on its own floor, and promote or discard the adapter on evidence rather than on a benchmark from elsewhere.
KEEP READING