Sembrick

The small language model platform

Your knowledge.
Your small model.

Turn specialist knowledge into a focused model. We’re building a place to prepare your data, choose an open model and train for the behaviour your task needs.

Request early access For users and developers.
One task. A model that understands it.

Support triage

An illustrative model project

Your example
“I’ve been charged twice for the same order.”
The behaviour you want
Billing Duplicate charge

Your examples define the task.
Independent evaluation measures the result.

In private development. Model Studio is being built and is not open to the public. Training and compute execution are still to come.

From task to model

Build one step at a time.

Data, models and evaluation belong in one considered workflow. This is the approach we’re building into Sembrick.

  1. 01

    Prepare the data.

    Inspect examples, define the behaviour and build a dataset with quality, provenance and independent review in mind.

  2. 02

    Choose and adapt.

    Select an open model and develop a fine-tuning recipe suited to your task and the resources available.

  3. 03

    Measure the result.

    Define success, compare against the base model and test behaviour on examples kept separate from training.

Start where you are

A dataset, a few examples,
or an idea.

Different starting points need different techniques. Sembrick is designed around all three.

Use an existing dataset
Understand the structure, duplicates and source groups in the examples you already have.
Grow a small dataset
Use seed examples to create varied candidates, then check quality before they enter training.
Create from a task
Define the desired behaviour first, then develop the examples a model should learn from.

Infrastructure · planned

A place for the whole build.

Data preparation, training jobs and model artifacts need a dependable environment. Sembrick’s infrastructure layer is planned to support that workflow.

We’re shaping it around real tasks and practical constraints. Hosting options, connected compute and job execution are still being defined.

Help shape Sembrick

What would you build?

Tell us about your task, the data you have and where you’d like a small model to help.

Email us for early access

Opens your email app with both team contacts. Please send a brief description, without confidential data or attachments. Early access is by invitation.