LLM Evaluation • Page 3

LLM Evals: What Exactly Are They?

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.

1. Definition of LLM Evaluations

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:

What exactly are LLM Evals?

LLM Evals can be defined as:

“LLM Evals are systematic, repeatable tests used to judge an LLM and LLM-powered systems against a clear criteria.”

Basically, LLM evaluations are tests that we apply either to:

There are three main aspects or characteristics:

1. Systematic
Tests should follow a proper, structured process.
2. Repeatable
The same evaluation should be runnable again on another version.
3. Clear Criteria
You must clearly define what you want to evaluate.

2. The Three Core Characteristics

2.1 Systematic

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.

Campus X Chatbot Example

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.

Systematic evaluation = structured data + broad coverage + deliberate testing.

2.2 Repeatable

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.

VersionSame Evaluation DatasetResult
Version 1YesPerformance of V1
Version 2YesPerformance 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.

Same evaluation setup + different system versions → meaningful comparison.

2.3 Clear Criteria

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:

  1. The answer should be correct.
  2. The answer should contain a simple explanation.
  3. The explanation should come from the course content.
  4. The answer should be safe.
  5. It should not contain unsafe content.
  6. It should not contain abusive language.
  7. It should not use a threatening tone.

You define these criteria and then evaluate the chatbot against them.

Without criteria: You are essentially doing vibe testing.
With clear criteria: You are doing proper evaluation.

3. LLM Eval Is NOT Just a Metric

This is a very important clarification.

If you come from Machine Learning or Deep Learning, you may initially think:

Evaluation = Metrics such as Accuracy, Precision, Recall, F1, etc.

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.

LLM Eval is not just a metric. It is the complete testing setup.

What is included in the complete testing setup?

All these things together form the evaluation setup.

RAG Example

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.

LLM Evals = the entire testing setup, not merely the metric.

4. What Questions Do LLM Evals Actually Answer?

The goal of an LLM eval is not simply to produce a score.

The goal is to answer practical engineering questions.

Model Selection
Can the model be used for a particular task/application?
Production Readiness
Is the system good enough to ship?
Prompt Optimization
Did Prompt V2 improve over Prompt V1?
RAG Groundedness
Is the RAG answer grounded in retrieved context?
Agent Performance
Is the agent completing the task correctly?
Safety
Is the chatbot safe for real users?
Performance
Is latency under control?

Therefore, evaluation gives us answers that can directly influence engineering decisions.

LLM Evals turn test results into practical product and engineering decisions.

5. Two Types of LLM Evals

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:

LLM EVALUATIONS | +-------------+-------------+ | | MODEL EVALS APPLICATION EVALS | | Evaluate the LLM Evaluate the itself LLM application

Model Evals

Model Evals are evaluations used to evaluate LLMs themselves.

Application Evals

Application Evals are evaluations used to evaluate LLM-based applications.

Important terminology disclaimer: The terms “Model Evals” and “Application Evals” are being used here as a conceptual simplification. They are not necessarily official standardized industry terms. In industry, both may simply be referred to as “LLM Evals,” with the context making it clear whether the model or the application is being evaluated.

6. Model Evals

Model Evals evaluate the model itself.

Model Evals = tests and evaluations of the capabilities of an LLM.

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.

6.1 Eight Major Capability Categories

The lecture identifies eight major capability categories for evaluating modern LLMs.

#CapabilityWhat Are We Asking?
1ReasoningCan the LLM reason through a problem and solve it step by step?
2KnowledgeDoes the LLM have relevant world/general knowledge?
3Basic MathematicsCan the LLM solve mathematical problems?
4CodingCan the LLM understand and generate code?
5Instruction FollowingCan the model follow multiple instructions and constraints correctly?
6Long-Context HandlingCan the LLM find and use the correct information inside a very large context?
7Multimodal UnderstandingCan the LLM understand modalities such as text, images, and audio?
8Tool UseCan the LLM properly utilize external tools?

6.2 Reasoning

The first capability is reasoning.

The question is:

Can the LLM reason through a problem step by step and solve it correctly?

6.3 Knowledge

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.

6.4 Basic Mathematics

The third capability is basic mathematics.

The question is:

Can the LLM solve mathematical problems correctly?

6.5 Coding

The fourth capability is coding.

The question is:

Can the LLM understand, generate, modify, and solve coding problems?

6.6 Instruction Following

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?

6.7 Long-Context Handling

The next capability is long-context handling.

The question is:

Can the LLM correctly find and use information from a very large context?

6.8 Multimodal Understanding

Another capability is multimodal understanding.

Depending on the model, this can involve understanding:

It can also involve producing appropriate outputs based on multimodal inputs.

6.9 Tool Use

The final capability is tool use.

The question is:

Can the LLM properly utilize external tools?

Examples include APIs, function calling, search, calculators, code execution, databases, and other external systems.

6.10 Benchmarks

These capabilities are evaluated using benchmarks.

Some examples mentioned in the lecture are:

Capability / AreaExample BenchmarkPurpose
General Knowledge / ReasoningMMLUTests questions across many subjects such as science, history, law, medicine, and more.
MathematicsGSM8KTests grade-school mathematical word problems.
Coding / Software EngineeringSWE-benchEvaluates software engineering/code-related problem solving.
Instruction FollowingIFEvalEvaluates whether models follow specified instructions and constraints.
Long ContextNeedle in a HaystackTests whether relevant information can be found inside a large context.
MultimodalMMMUEvaluates multimodal understanding across diverse tasks.

6.11 How to Think About Benchmarks

Capability | +---- Reasoning | +---- Knowledge | +---- Mathematics | +---- Coding | +---- Instruction Following | +---- Long Context | +---- Multimodal | +---- Tool Use | v Benchmarks | v Model Results | v Capability Comparison

6.12 Why AI Engineers Need Model-Eval Literacy

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.

You may never personally perform frontier-model evaluation, but you should understand model evaluations and know how to read benchmark results.

7. Application Evals

Now we come to the second category:

Application Evals

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.

7.1 Why Application Evals Exist

LLM applications are not just the LLM.

Especially for beginners, there is a tendency to think:

“If the LLM is the brain, then everything about the application must be about the LLM.”

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.

7.2 Components Around the LLM

An LLM-powered application may include:

Therefore, in a proper production-grade LLM application:

The LLM is just one component.

7.3 RAG Example

Suppose you are building a RAG system.

You may have:

RAG APPLICATION | +----------------+----------------+ | | | Retrieval LLM Application | | | Embeddings Prompt Guardrails Vector DB Generation Monitoring Reranker Feedback

The retrieval system itself may have multiple components:

Every one of these can affect the final quality of the application.

8. Smartphone Analogy

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:

Does having a powerful processor guarantee that the smartphone is a good smartphone?

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.

Powerful Processor | v Does NOT automatically mean | v Excellent Smartphone Powerful LLM | v Does NOT automatically mean | v Excellent AI Application

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.

9. Definition of Application Evals

Application Evals assess the behavior and performance of an LLM-powered application, either at the level of the entire system or at the level of a specific component within it.

This means application evaluation can happen at:

Component Level

Entire System Level

10. RAG Application: Component-Level Evaluation

Suppose you built a RAG chatbot.

You can evaluate the entire chatbot:

But you can also evaluate individual components.

For example:

Retriever
Is the correct information being retrieved?
Embedding Model
Does it represent semantic meaning effectively?
Reranker
Does it correctly reorder retrieved documents?
Prompt
Does the prompt produce the desired behavior?

This is why application evaluation is broader than simply evaluating the LLM.

11. The Central Question: Model vs Application

Model Evaluation asks:

“Can the model do this?”

For example:

Application Evaluation asks:

“Will our product/system work correctly?”

For a Campus X chatbot, application evaluation can answer:

12. Campus X Chatbot — Complete Example

Suppose we build a Campus X educational chatbot.

Model Evaluation

We might ask:

  • Can the underlying model reason?
  • Can it follow instructions?
  • Can it solve mathematics?
  • Can it handle long context?
  • Can it use tools?

Application Evaluation

We might ask:

  • Did the chatbot answer the student's question correctly?
  • Did it use the correct course material?
  • Was the response grounded?
  • Was the response easy for a beginner to understand?
  • Did hallucination occur?
  • Was the response safe?
  • Was the response fast enough?
  • Was the application reliable?

13. Why Application Evals Are the Main Topic for AI Engineers

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.

Model evaluation tells us how capable the underlying model is. Application evaluation tells us whether the actual product we built works correctly.

14. Practical Focus of the Course

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.

15. Important Rule When Watching Future LLM-Evaluation Content

When you watch future content on YouTube or elsewhere that says:

“LLM Evaluation”

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.

16. Complete Lecture Summary

Let's summarize the entire flow covered in this lecture.

Part 1 — WHY

We first discussed:

Part 2 — WHAT

We then discussed:

Part 3 — TYPES

We divided LLM Evals conceptually into:

TypePurpose
Model EvalsEvaluate the LLM itself and its capabilities.
Application EvalsEvaluate the LLM-powered application/system.

Part 4 — Model Evals

We learned that Model Evals evaluate the capabilities of an LLM.

The eight capability categories discussed were:

  1. Reasoning
  2. Knowledge
  3. Basic Mathematics
  4. Coding
  5. Instruction Following
  6. Long-Context Handling
  7. Multimodal Understanding
  8. Tool Use

We also saw benchmark examples including:

Part 5 — AI Engineer Relevance

We learned that an AI Engineer may not frequently perform frontier-model evaluations personally.

However, an AI Engineer should understand:

Part 6 — Application Evals

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.

Part 7 — Component + System Evaluation

Application Evals can happen at:

Part 8 — The Core Difference

MODEL EVAL | +--> "Can the model do X?" | +--> Evaluate model capabilities | +--> Benchmarks | +--> Model-level comparison APPLICATION EVAL | +--> "Does my product/system work correctly?" | +--> Evaluate application behavior | +--> Component + end-to-end evaluation | +--> Production readiness

17. Final Mental Model

Remember These Four Lines

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.

For an AI Engineer: Model Evals give you model literacy; Application Evals give you practical engineering control over the system you build.

18. Transition to the Next Topic — HOW

So far, we have covered:

WHY
Why do we need LLM evaluations?
WHAT
What exactly are LLM Evals?
TYPES
Model Evals and Application Evals.

Now we move toward:

HOW are LLM evaluations actually performed?

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.

End of Page 3

Next: How Application-Level LLM Evaluations are actually performed — datasets, evaluation criteria, evaluators, metrics, workflows, and practical implementation.