Glossary
Fine-tuning
Fine-tuning is extra training of an existing model on your own labelled examples so that it specialises in your task, usually becoming more accurate.
Swiss-hosted inference. Nothing you send is ever stored.Swiss data residency
What Fine-tuning means
Fine-tuning takes a model that has already been pre-trained on broad data and continues training it on a narrower, labelled dataset for a specific task or domain. It can update all weights (full fine-tuning) or only a small subset (adapters, LoRA, a classification head). For encoders it is the classic recipe: pre-train with a masked language model objective, then fine-tune on labelled examples.
Fine-tuning usually beats zero-shot and in-context few-shot approaches on the target distribution, at the cost of needing labelled data, a training run and a model to maintain. A classic fine-tuned BERT classifier has a fixed label set; changing the labels means retraining.
Fine-tuning Laya
Laya is itself a fine-tuned encoder (ModernBERT-large or mmBERT-base, fully fine-tuned with RLCD), and it can be fine-tuned further. The best evidence is on the model card's typed-decisions benchmark (400 cases, 2,000 decisions, four workflows):
| model | accuracy | Brier |
|---|---|---|
| laya-typed-decisions (fine-tuned) | 0.766 | 0.062 |
| laya (base) | 0.362 | 0.316 |
| majority class | 0.461 |
The fine-tuned checkpoint wins all four workflows: invoice processing 0.804, security incidents 0.766, customer service 0.764, agent-trace observability 0.730. The base checkpoints sit below the majority-class baseline on this benchmark, which is the honest reason to fine-tune: specialised decisions need specialised training.
Unlike a classic classifier, a fine-tuned Laya still accepts request-time options, so you keep the flexibility of typed questions. After fine-tuning, refit temperatures on held-out data. See typed decisions and Laya vs fine-tuned BERT.
Related terms
How Fine-tuning connects to the rest of the vocabulary.