Training foundation models · Question 11 of 12
A pharmaceutical company has 2 million unlabelled specialist articles and 3,000 reviewed question-answer pairs. The model should understand the specialist language and answer questions in the company format. Which sequence makes sense?
1 answer is correct. Hard
Show answer
Answer
- A First continued pre-training with the articles, then fine-tuning with the pairs
Explanation
Unlabelled domain texts are suited to continued pre-training, which conveys the specialist language. The labelled pairs are then used for fine-tuning on format and task. New pre-training with only 3,000 examples would be completely insufficient.