Research areas

Overview

Research areas

Model adaptation

A general-purpose model is rarely the cheapest or fastest way to do one specific job well. We adapt open models to your domain and constraints, and only when the evaluation shows it beats a well-prompted general model.

  1. baseline

    prompted model

  2. data

    curate · label

  3. adapt

    fine-tune · distil

  4. evaluate

    vs baseline

  5. serve

    quantise · deploy

Fig. 1Adaptation is only worth shipping when it beats the baseline on your evals

Typical problems

  • API costs grow faster than usage justifies
  • Latency is too high for the product experience
  • Data cannot leave your infrastructure, so hosted APIs are not an option
  • A general model keeps missing the terminology or format of your domain

What you get

  • A baseline comparison with general-purpose models on your tasks
  • A curated training dataset and a reproducible training pipeline
  • An adapted model with measured quality, latency and cost
  • Deployment on your infrastructure, with serving optimised for your hardware

How we approach it

Every adaptation project starts with a baseline: the best result a general model achieves with good prompting and retrieval. Fine-tuning has to beat that on your evaluation set, or it isn’t worth the extra complexity.

Most of the work is in the data. We curate and clean training examples from your domain, keep a held-out test set the model never sees, and track quality, latency and cost side by side, so the trade-offs are clear before anything is deployed.