Small language and open-weight model training
Not every job needs a frontier model. We train and fine-tune small or open-weight models for the specific task you have, then run them locally, so you keep control of the model, the data and the cost.
Train or fine-tune small and open-weight models for one job, such as classification or customer care, and run them where your data lives.
- Open-weight language models
- Small language models
- Vision-language models
- Embedding models
- Supervised fine-tuning
- Parameter-efficient adapters
- Prompt and few-shot baselines
- Ollama
- OpenAI-compatible endpoints
- Docker containers
- GPU or CPU hosts
The challenge
Many enterprise tasks are narrow: sort a document, answer a customer question, extract a field. Sending every one of them to a large hosted model is expensive, slower than it needs to be, and moves sensitive data outside your boundary.
Teams that try to train their own model often stall on the same things: unlabelled data, no way to measure quality, and no plan for running the result. A model that cannot be evaluated, governed or hosted never reaches production.
What it does
Training data preparation
Collect, clean, label and split your examples, with sensitive fields removed or masked first, so the training set is representative and safe to use.
Task-specific fine-tuning
Adapt a small or open-weight model to one defined job, such as document classification or customer-care answers, using your own labelled examples rather than a generic prompt.
Model comparison and selection
Run candidate open-weight models against the same test set, and compare quality, speed and hardware needs before you commit to one.
Evaluation and retraining plan
Define a fixed test set and acceptance thresholds, and set up a repeatable path to retrain when your documents, products or policies change.
Local and private deployment
Package the trained model to run on your own servers or workstations, so documents and conversations never leave your environment.
Integration and human hand-off
Serve the trained model through a standard API so your applications, including chat assistants and document pipelines, can call it, with a route to a person when confidence is low.
How it works
Define the job
Agree the single task, the quality bar, the data available, the hardware you can run on and the limits on what the model may see.
Prepare the data
Gather and label examples, mask sensitive fields, and hold back a test set that the model never sees during training.
Train and evaluate
Fine-tune candidate models, score each against the held-out test set, and review the errors with your subject experts before choosing.
Deploy and monitor
Run the chosen model locally behind the harness, track quality over time, and retrain when the measured results begin to drift.
Under the hood
- Model families
- Open-weight language models
- Small language models
- Vision-language models
- Embedding models
- Training methods
- Supervised fine-tuning
- Parameter-efficient adapters
- Prompt and few-shot baselines
- Local serving
- Ollama
- OpenAI-compatible endpoints
- Docker containers
- GPU or CPU hosts
- Evaluation
- Held-out test sets
- Replay of past documents
- Side-by-side model comparison
- Human review of errors
Business outcome
A model sized to the job. You use a smaller model where it is enough, and keep larger models for the work that truly needs them, which keeps running costs predictable.
Data stays inside your boundary. Documents and conversations are processed on your own infrastructure, which supports data residency and privacy requirements without sending content to an outside service.
Quality you can measure. Each model comes with test results on your own data, so a decision to deploy rests on evidence and not on a demo.
No lock-in to one vendor. Because the model is open-weight and runs on standard serving tools, you can move it, replace it or retrain it as better models appear.
Where it is used
- Training a model to classify and route mixed incoming documents into your own catalogue of document types.
- Training a customer-care model on your own policies and past answers, with hand-off to a person for difficult cases.
- Running an extraction or summarisation model on a local server where documents cannot leave the premises.
- Comparing several open-weight models on your own test set before choosing one for production.
Frequently asked questions
Do we need a large dataset?
Usually not. Narrow tasks often work with a modest set of well-labelled examples. We assess your data in the first step and tell you honestly whether it is enough before training starts.
Can the model run without internet access?
Yes. The trained model can be served on your own hardware, including machines with no outbound connection, using standard local serving tools. Hardware needs depend on the model size we choose together.
Which base models do you use?
We stay vendor-neutral. We compare open-weight and small models from several families against your test set and recommend the one that fits your quality, speed and hardware limits.
Is training the only option?
No. Sometimes a well-prompted model with retrieval is enough and training adds nothing. We test that baseline first, and only recommend fine-tuning when the measured gap justifies it.
See it on your own documents and data
Tell us what you want to automate and we will show you how it runs inside the harness.