A foundation model goes through a recognizable lifecycle, from choosing its data to learning from users after launch. Most organizations only work in the later stages, because they start from a model someone else has already pre-trained. This lesson walks through each stage and shows which ones a typical business actually touches.
Data selection and model selection
For a model builder, the lifecycle starts with data selection: gathering a huge, broad and high-quality dataset, because the model can only learn what its data contains. Gaps or bias in the data become gaps or bias in the model.
For most businesses, the practical starting point is model selection: choosing an existing foundation model that fits the job. Factors include the kind of input and output it handles, accuracy on the task, context window size, speed, price and data-usage terms. AWS's Well-Architected Generative AI Lens places a scoping phase before this, to confirm the business problem really needs generative AI and how success will be measured.
Pre-training
Pre-training is where a foundation model learns general knowledge and language patterns by processing enormous amounts of mostly unlabeled data, for example by repeatedly predicting the next token. It takes vast computing power, weeks or months of time and a very large budget, which is why few organizations do it themselves. The output is a general-purpose model that can do many things reasonably well.
Fine-tuning
Fine-tuning continues training a pre-trained model on a smaller, task-specific or domain-specific dataset, often labeled. It adapts the model's style, format or specialist knowledge, for example to write in a company's tone or handle legal language. It is far cheaper than pre-training, but it still costs money and needs good data, so prompt engineering and retrieval are usually tried first. Domain 3 covers the methods.
Evaluation and deployment
Evaluation tests the model before release: with benchmark datasets, automatic metrics, or human reviewers judging the quality, accuracy and safety of its answers. It should check the results against the success measures defined at the start, not just general benchmarks.
Deployment makes the model available to users, usually through an API inside an application. On AWS, that might mean calling a model through Amazon Bedrock, or hosting it on an Amazon SageMaker AI endpoint.
Feedback
After launch, the cycle continues. Teams monitor outputs, collect user feedback such as ratings or corrections, and watch metrics like accuracy and harmful content. That feedback guides the next round of prompt changes, data updates, fine-tuning or a switch to a better model. A deployed model that is never revisited slowly falls behind changing needs and data.
| Stage | Typical business role |
|---|---|
| Data selection, pre-training | Done by the model provider |
| Model selection | Choose from existing FMs |
| Fine-tuning | Optional, when prompts and retrieval aren't enough |
| Evaluation, deployment, feedback | Always the business's job |
On the exam
Expect questions that describe an activity and ask which stage it belongs to. "Collecting thumbs-up and thumbs-down ratings" is feedback. "Training on a company's own labeled examples" is fine-tuning. "Learning from a huge general dataset" is pre-training. Also remember the order: pre-training comes before fine-tuning, and evaluation comes before deployment.