Model Training

Fine-tuning is worth doing when prompting has genuinely run out of room, which is later than most people assume. A better prompt, better retrieval, or a cleaner set of examples solves the majority of cases at a fraction of the cost and without freezing a model version into your product.

When training is the right answer, the dataset is the project. Collecting the examples, deciding what a correct output looks like, and getting agreement between the people labeling it take longer than the training run and determine whether it works.

Evaluation comes before training, not after. A scored set of questions with known right answers is what tells you the fine-tune helped, and it is the same set you re-run when the base model changes underneath you, which it will.

// what you get

What's included.

  • An honest assessment of whether fine-tuning beats prompting and retrieval for your case
  • Dataset construction: collection, labeling guidelines, and inter-rater agreement
  • An evaluation set built before training, with the scoring criteria written down
  • Fine-tuning, LoRA adapters, and distillation to a smaller model where cost or latency matters
  • Versioning of models, datasets, and prompts, so a regression can be traced to a change
  • Re-evaluation when the base model updates, rather than assuming the numbers hold
// where it fits

The situations this is usually bought for.

A model that will not follow a house format

Output that is right but shaped wrong, every time. A few hundred good examples usually fix the format when prompting has stopped helping.

Domain language a general model gets wrong

Industry terms read as ordinary words. Training on real documents teaches the vocabulary and the conventions around it.

A capable model too slow or too expensive to run

Distilling to a smaller model that keeps the accuracy on your specific task, and measuring what was lost rather than hoping.

Light tracing connections across a dark surface
// say hello

Tell us what you are trying to fix.

Discovery calls are free and usually last 30 minutes. We listen first, and we will tell you honestly if model training is the wrong thing to spend money on.