Fine-Tuning a Language Model on Question-Answer Pairs
Problem statement
How would you fine-tune a language model to answer support questions in the house style? Agents wrote several question paraphrases and reference answers for each knowledge-base article. Training must teach the model when to stop, and evaluation must show whether it handles real questions beyond familiar article paraphrases.
Explain training examples, evaluation, optimization, and deployment with monitoring and rollback.
Show hint
What exactly does the model see, and get a loss on, after the last token of an answer during training?
Relevant roles: ML Engineer, LLM Engineer, AI Engineer and Gen AI Engineer.