Exam-style
A team built a customer-facing help assistant on a foundation model from the Amazon Bedrock catalogue. It shipped a month ago. Under every answer sit thumbs-up and thumbs-down buttons, and the team stores each rating with the prompt and answer that produced it. One answer in twenty gets a thumbs-down, most of them refund questions where the assistant quotes a policy that changed last year. Which stage of the FM lifecycle is this, and what does the signal feed?
Reveal the answer
C. Feedback, and it feeds evaluation first, then fine-tuning when the ratings show a gap that is consistent rather than scattered
The stages run data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback. Collecting what real users think of real answers is feedback, the last stage and the one that never finishes.
Evaluation is where the ratings go first: they become a scored set of real prompts, and every later prompt or model change gets measured against it rather than a demo. Fine-tuning comes second, and only when the gap is systematic. Here it is not. The refund answers are wrong because the policy changed and the model is drawing on the old one, which a corrected help-centre article fixes in an afternoon.
The other four misfile the activity. Data selection is choosing the corpus a model learns from, before training rather than after launch. Deployment covers the assistant running, not the ratings coming back. Pre-training is the long run that turns a broad corpus into general-purpose weights, and almost no team runs one: a foundation model arrives pre-trained on hundreds of billions of tokens. Model selection is a real stage and the ratings do inform it, but swapping the model responds to the measurement rather than naming it.
Q. A live assistant is collecting thumbs-up and thumbs-down from users, stored with the prompt and the answer. Which stage of the FM lifecycle is that, and what does it feed?
A. Feedback. It feeds evaluation first, and fine-tuning after that, when the ratings show a consistent gap rather than scattered complaints.
Why? Learn the seven in order: data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback. Feedback is the one that runs forever, because it starts the day the thing goes live and never finishes. Ratings on real answers, stored next to the prompts that produced them, become a scored set you can re-run whenever anything changes, which is why evaluation is the first place they go. Fine-tuning is the next place, and it is the right move only when the ratings cluster: one topic, one tone, one shape of answer, wrong over and over. Scattered thumbs-downs usually mean a prompt or a stale document. Fix those before you reach for customisation. The tempting misfile is pre-training, because ratings feel like training data. Almost nobody runs pre-training. A foundation model arrives with that stage already done by whoever built it, and the stages a team actually owns start at model selection.