Model adaptation adjusts a pretrained generative AI model to improve its output for a specific task or domain. It reuses knowledge from general training rather than building a model from scratch, making it useful for enterprise applications that require specialized vocabulary or business context. section 9.1 section 9.0
Two ways to frame the target are:
- Domain adaptation moves a model to a related domain, such as from general news to medical news. section 9.1
- Task adaptation moves a model to a related task, such as from sentiment analysis to emotion detection. section 9.1
Fine-tuning continues training on a smaller, specialized dataset. Full fine-tuning updates all model parameters; parameter-efficient fine-tuning (PEFT) updates fewer. Low-rank adaptation (LoRA), a PEFT technique, introduces trainable low-rank matrices while keeping most pretrained parameters fixed. The choice depends on the required adaptation, model size and available compute. section 9.1.1 section 9.6
Adapting for medical questions
A model trained on general articles may perform poorly on medical questions. Fine-tuning it on a smaller, relevant dataset can help it learn patterns and vocabulary for that domain. The dataset should be clean, consistent and representative of the intended task; performance still needs evaluation. section 9.1 section 9.1.2 section 9.8
Start with a clear use case, not fine-tuning by default. An existing SaaS copilot may be the best starting point; where it is unsuitable, begin with prompt engineering and consider retrieval-augmented generation (RAG). These techniques can complement fine-tuning rather than replace one another. Fine-tuning adds data and compute costs, can cause catastrophic forgetting, and requires task-specific metrics and often several evaluation cycles before production. section 9.0 section 9.8 section 9.6






