← AI Terminology
Domain Adaptation
Domain adaptation adapts a model trained on a source distribution to perform well on a different but related target distribution, often with limited target labels.
Classic ML problem revived constantly in industry NLP and vision.
Classic ML problem revived constantly in industry NLP and vision.
Why It Matters in AI
General web models underperform on legal, medical, or stack-internal text. Adaptation (continued pretrain, finetune, adapters, retrieval) is how teams specialise foundation models affordably.
Key Points
| Aspect | Description |
|---|---|
| LLM | Domain corpora mid-training; RAG for knowledge |
| Eval | Target-domain golden sets |
| Risk | Catastrophic forgetting of general skills |
| Methods | Fine-tune, continued pretrain, adversarial DA, adapters |
| Related | Transfer learning, continual learning |
| Settings | Unsupervised, semi-supervised, few-shot target labels |
Simple Analogy
A chef trained in one regional cuisine spending a week learning local ingredients to cook well in a new city — same skills, new distribution.
Common Usage Examples
- Continued pretrain on domain corpus
- LoRA on support tickets
- Measure domain perplexity drop
- Hybrid: adapt + retrieve private docs
Summary
In short: Domain adaptation specialises models from source data to a different target distribution — how general models become useful in niche settings.