Complete lecture notes — full content preserved and organized for study.
This page covers what LLM evaluations are, their three defining characteristics, why an eval is more than a metric, the complete testing setup, practical questions answered by evaluations, Model Evals vs Application Evals, model capabilities and benchmarks, and why Application Evals are especially important for AI Engineers.
Now that we understand why LLM evaluations are needed, why they are important, and how LLM evaluations are different from traditional software testing, we can move to the main topic:
LLM Evals can be defined as:
Basically, LLM evaluations are tests that we apply either to:
There are three main aspects or characteristics:
Systematic means that you are not doing vibe testing.
Suppose five questions come to your mind. You ask those five questions, receive answers, and feel that everything looks correct. That is not proper LLM evaluation.
Instead, what you do is create proper datasets. In those datasets, you try to cover different kinds of edge cases so that you can properly test your chatbot or LLM-based system.
Suppose I am building a Campus X chatbot.
What I can do is randomly take around 100 real-user conversations with my system. I can create a database or dataset from those conversations and then perform testing on that dataset.
This helps me understand the real-world behavior of my chatbot.
Repeatable means that tomorrow, if you change something in your system, you should still be able to evaluate the system in exactly the same way.
You may change:
Even after all these changes, you should be able to apply the same evaluation setup.
Suppose you have a test dataset. You run Version 1 against it and get a certain performance. You then run Version 2 against the same dataset.
| Version | Same Evaluation Dataset | Result |
|---|---|---|
| Version 1 | Yes | Performance of V1 |
| Version 2 | Yes | Performance of V2 |
Now you can compare whether Version 1 or Version 2 is better.
This is why repeatability is extremely important. The dataset and evaluation process need to remain sufficiently consistent so that you can determine whether the system is actually improving.
The third and most important characteristic is that LLM evaluations depend on what criteria you want to evaluate against.
Suppose you are building a Campus X chatbot. There can be many criteria:
You define these criteria and then evaluate the chatbot against them.
This is a very important clarification.
If you come from Machine Learning or Deep Learning, you may initially think:
In traditional ML, when someone asks how we evaluate a model, we often talk about metrics such as:
Therefore, it is natural to think that LLM Evals are simply a set of metrics.
But that is not the correct mental model.
All these things together form the evaluation setup.
Suppose you have a RAG chatbot and you want to evaluate the retriever.
The retriever becomes one component of your LLM evaluation setup.
You might evaluate it using a criterion such as retrieval accuracy.
The dataset you prepared for evaluation is also part of the evaluation setup.
Whether you run the evaluation offline or after deployment is also part of the setup.
If you use a tool such as Ragas because you are evaluating a RAG application, that tool is also part of your evaluation workflow.
The goal of an LLM eval is not simply to produce a score.
The goal is to answer practical engineering questions.
Therefore, evaluation gives us answers that can directly influence engineering decisions.
Before discussing exactly how LLM evaluations work, there is an important distinction that should be clear in your mind.
Conceptually, LLM Evals can be divided into two parts:
Model Evals are evaluations used to evaluate LLMs themselves.
Application Evals are evaluations used to evaluate LLM-based applications.
Model Evals evaluate the model itself.
The main idea is to test and evaluate the capabilities of a model.
When a new LLM is released, we want to know:
You may have seen new LLM releases where people say:
“This model ranked at the top on a particular benchmark or leaderboard.”
What is happening there?
There are evaluation datasets and benchmarks designed to measure particular capabilities. When a new model arrives, the model is tested against those benchmarks. The results are then documented and often published so that people can understand how capable the model is.
The lecture identifies eight major capability categories for evaluating modern LLMs.
| # | Capability | What Are We Asking? |
|---|---|---|
| 1 | Reasoning | Can the LLM reason through a problem and solve it step by step? |
| 2 | Knowledge | Does the LLM have relevant world/general knowledge? |
| 3 | Basic Mathematics | Can the LLM solve mathematical problems? |
| 4 | Coding | Can the LLM understand and generate code? |
| 5 | Instruction Following | Can the model follow multiple instructions and constraints correctly? |
| 6 | Long-Context Handling | Can the LLM find and use the correct information inside a very large context? |
| 7 | Multimodal Understanding | Can the LLM understand modalities such as text, images, and audio? |
| 8 | Tool Use | Can the LLM properly utilize external tools? |
The first capability is reasoning.
The question is:
The second capability is knowledge.
We want to know whether the model has basic world knowledge and general knowledge.
This is also related to the concept of a knowledge cutoff date. The idea is that information available before the model's knowledge cutoff may be part of its learned knowledge, while information created after that cutoff may not be present in its pretrained knowledge.
The third capability is basic mathematics.
The question is:
The fourth capability is coding.
The question is:
The fifth capability is instruction following.
Suppose you give an LLM ten instructions. Can it follow all ten instructions correctly, one after another, while respecting the constraints?
The next capability is long-context handling.
The question is:
Another capability is multimodal understanding.
Depending on the model, this can involve understanding:
It can also involve producing appropriate outputs based on multimodal inputs.
The final capability is tool use.
The question is:
Examples include APIs, function calling, search, calculators, code execution, databases, and other external systems.
These capabilities are evaluated using benchmarks.
Some examples mentioned in the lecture are:
| Capability / Area | Example Benchmark | Purpose |
|---|---|---|
| General Knowledge / Reasoning | MMLU | Tests questions across many subjects such as science, history, law, medicine, and more. |
| Mathematics | GSM8K | Tests grade-school mathematical word problems. |
| Coding / Software Engineering | SWE-bench | Evaluates software engineering/code-related problem solving. |
| Instruction Following | IFEval | Evaluates whether models follow specified instructions and constraints. |
| Long Context | Needle in a Haystack | Tests whether relevant information can be found inside a large context. |
| Multimodal | MMMU | Evaluates multimodal understanding across diverse tasks. |
If you are going to become an AI Engineer, you probably will not spend most of your time performing frontier-level model evaluations.
Why?
Imagine a new LLM is released. Evaluating that new model across many benchmarks, documenting the results, and publishing those evaluations is generally the work of large frontier labs and model developers.
Your practical responsibility is more likely to be:
For example, when starting a project, you may need to decide:
Model evaluation knowledge can help you make these decisions.
Now we come to the second category:
This is the category that is especially important for an AI Engineer.
Your work as an AI Engineer is to build LLM-based applications. Therefore, evaluating those applications is also your responsibility.
Application Evals are the main practical focus of this topic.
LLM applications are not just the LLM.
Especially for beginners, there is a tendency to think:
But that is not true.
As you gain experience in AI Engineering and start building larger applications, you realize that although the LLM is very important, many other components are required for the complete application to work correctly.
An LLM-powered application may include:
Therefore, in a proper production-grade LLM application:
Suppose you are building a RAG system.
You may have:
The retrieval system itself may have multiple components:
Every one of these can affect the final quality of the application.
A very useful analogy is a smartphone.
A smartphone may have a processor such as a Snapdragon or MediaTek chip.
Chip manufacturers often publish benchmark scores to demonstrate how powerful their processors are.
These benchmarks tell us something about the capability of the processor.
But ask yourself:
Obviously, no.
A smartphone also requires:
All of these components need to work properly together.
Similarly, if you have a powerful LLM, that does not automatically guarantee that your LLM application will be good.
Just as you evaluate the battery, display, camera, processor, operating system, etc. of a smartphone, you need to evaluate the different components of an LLM application.
The model evaluation is often performed by the model developers/frontier labs.
But the evaluation of the complete system that you build around the model is your responsibility as an AI Engineer.
This means application evaluation can happen at:
Suppose you built a RAG chatbot.
You can evaluate the entire chatbot:
But you can also evaluate individual components.
For example:
This is why application evaluation is broader than simply evaluating the LLM.
For example:
For a Campus X chatbot, application evaluation can answer:
We might ask:
We might ask:
As an AI Engineer, your job is usually not to create the next frontier LLM.
Your job is more likely to be building applications using existing LLMs.
Examples include:
Therefore, you need to know whether the system you built is actually working.
This makes Application Evals extremely important.
The course will focus much more heavily on Application Evals.
There is still value in understanding Model Evals because:
However, you may never personally need to perform large-scale frontier-model benchmarking.
Application evaluation is much more likely to be part of your day-to-day work as an AI Engineer.
When you watch future content on YouTube or elsewhere that says:
you should understand that the term can refer to both model-level and application-level evaluation.
In practical AI engineering content, a large amount of evaluation discussion is about evaluating the LLM-powered application, rather than performing frontier-model benchmark research.
Always look at the context to determine which kind of evaluation is being discussed.
Let's summarize the entire flow covered in this lecture.
We first discussed:
We then discussed:
We divided LLM Evals conceptually into:
| Type | Purpose |
|---|---|
| Model Evals | Evaluate the LLM itself and its capabilities. |
| Application Evals | Evaluate the LLM-powered application/system. |
We learned that Model Evals evaluate the capabilities of an LLM.
The eight capability categories discussed were:
We also saw benchmark examples including:
We learned that an AI Engineer may not frequently perform frontier-model evaluations personally.
However, an AI Engineer should understand:
We then moved to Application Evals, which are especially important for AI Engineers.
We learned that an LLM application is much more than the LLM itself.
It can contain:
Therefore, all these components can require evaluation.
Application Evals can happen at:
1. LLM Evals are systematic, repeatable tests against clear criteria.
2. LLM Evals are not just metrics — they are the complete testing setup.
3. Model Evals tell us how capable the underlying LLM is.
4. Application Evals tell us whether the LLM-powered product actually works.
So far, we have covered:
Now we move toward:
And there is one final important disclaimer:
The upcoming “HOW” discussion is primarily going to be taught from the perspective of Application Evals, not from the perspective of frontier-level Model Evals.
Next: How Application-Level LLM Evaluations are actually performed — datasets, evaluation criteria, evaluators, metrics, workflows, and practical implementation.