Question

A startup wants a foundation model to classify incoming support tickets into five categories using a consistent output format. They have only about 40 labeled examples and need to launch within a week with minimal upfront cost. The base model performs poorly with a simple instruction alone. Which approach best balances quality, cost, and time-to-launch for this situation?

Accepted Answer

Include a few representative labeled examples directly in the prompt to guide the model's classification behavior

Answer

Fine-tune the foundation model on the 40 labeled examples before deployment

Answer

Run continued pre-training on a large corpus of unlabeled support tickets

Answer

Build a Retrieval Augmented Generation pipeline with a vector database of past tickets