Fundamentals of GenAI · Topic 2 of 7
Foundation Model Lifecycle
Follow the lifecycle of a foundation model from picking a model through pre training, adapting, evaluating, deploying and collecting feedback.
- 4 min
- The stages are data selection, model selection, pre training, adaptation, evaluation, deployment and feedback.
- Pre training is expensive and normally done by the model provider.
- Adapt a model in rising order of effort: prompt engineering, RAG, fine tuning, continued pre training, training from scratch.
What it is
The foundation model lifecycle is the set of stages that a GenAI solution goes through, from choosing a model to running it and learning from users. It looks like the classic ML lifecycle, but most teams start from an existing model.
Understanding the stages helps you place a question: is it about picking a model, adapting it, testing it or running it?
How it works
The stages
| Stage | What happens |
|---|---|
| Data selection | Choose and prepare the data the model learns from or uses |
| Model selection | Pick a base model that fits cost, quality, speed and needs |
| Pre training | Train a large model on broad data, usually done by the model provider |
| Fine tuning or adaptation | Make the model fit your task or domain |
| Evaluation | Measure quality, safety and fit with business goals |
| Deployment | Make the model available to applications |
| Feedback and monitoring | Collect user feedback and track performance |
Pre training
Pre training builds the foundation model. It needs enormous data, a lot of computing power and a long time, so it is costly and most organisations do not do it. They use a model that a provider has already pre trained.
Choosing how to adapt
Once you have a base model, you can adapt it in increasing levels of effort.
| Approach | What it does | Cost and effort |
|---|---|---|
| Prompt engineering | Writes better instructions and examples | Lowest |
| RAG | Adds your documents to the prompt at request time | Low to medium |
| Fine tuning | Further trains the model on your labeled examples | Medium to high |
| Continued pre training | Trains on large unlabeled domain data | High |
| Train from scratch | Builds a new model | Highest |
A good rule is to start with the cheapest approach that meets the need and move up only if it does not. The details of these methods are in the foundation model application chapter.
Evaluation and deployment
Evaluation checks the answers against the goal, using human review, test sets and automatic scores. Deployment can use a managed service, such as Amazon Bedrock, where AWS hosts the model and you call it through an API. You can also host a model yourself on Amazon SageMaker AI when you need more control.
Feedback
After release, teams gather user ratings, review mistakes and watch cost and delay. Feedback can lead to better prompts, new documents for RAG, or a new round of fine tuning. This closes the loop back to earlier stages.
Responsible use throughout
Safety checks belong at every stage, not only at the end. Data should be reviewed for quality and bias, outputs should be tested for harmful content, and live use should be watched.
What to remember for the exam
- The stages are data selection, model selection, pre training, adaptation, evaluation, deployment and feedback.
- Pre training is expensive and normally done by the model provider.
- Adapt a model in rising order of effort: prompt engineering, RAG, fine tuning, continued pre training, training from scratch.
- Start with the simplest approach that meets the need.
- Evaluation covers quality, safety and business fit, not just a single score.
- Amazon Bedrock offers managed access to models through an API, and SageMaker AI offers more control.
- Feedback from users and monitoring feed improvements back into earlier stages.
- Safety and bias checks apply at every stage.
When a question asks for the lowest cost way to improve answers, think prompt engineering or RAG before fine tuning.
Common mistakes
Assuming each company trains its own foundation model. That is rare because of the cost. Most use a provider's pre trained model.
Jumping to fine tuning first. It needs labeled data and effort. Better prompts or RAG may solve the problem more cheaply.
Skipping evaluation before launch. A model that sounds fluent can still be wrong or unsafe.
Treating deployment as the final step. Feedback and monitoring continue, and they drive later changes.
Mixing up pre training and fine tuning. Pre training uses broad data to build the base model. Fine tuning adapts it to a narrower task.
Check yourself
Who usually performs pre training?
The model provider, because it needs huge data and computing power.
Put these in order of rising effort: fine tuning, prompt engineering, RAG.
Prompt engineering, then RAG, then fine tuning.
Why collect feedback after deployment?
It shows where answers fail and guides improvements to prompts, data or the model.
Topic 2 of 7 in Fundamentals of GenAI
Test yourself on Fundamentals of GenAI
The full AWS AIF-C01 question bank has 849 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by AWS. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.