A pharmaceutical company has millions of unlabeled internal research articles. It wants the foundation model to absorb this specialized vocabulary and writing patterns before any task-specific training, without supplying question-answer pairs. Which customization method does this?
Show answer and explanation
Answer: C
Continued pre-training runs further unsupervised training on a large, unlabeled domain corpus, so the model absorbs terminology and style. No labels are required, which suits this company's archive. Supervised fine-tuning requires labeled examples for a task. RAG and few-shot prompts leave the model weights unchanged.
Why the other options are wrong
- A. Fine-tuning needs labeled input-output pairs and teaches specific behaviors for a task. The scenario supplies only unlabeled articles, so this method does not fit.
- B. RAG retrieves documents at query time and places them into the prompt. The model's weights are unchanged, so it does not learn the domain's vocabulary.
- D. Few-shot prompts add examples to each request. They guide responses temporarily and do not update what the model has learned from its corpus.