PAGE 02 · LLM EVALUATIONS

What Exactly Are LLM Evals?

A detailed guide to understanding LLM evaluations, systematic and repeatable testing, evaluation criteria, model evaluations, application evaluations, benchmarks, and the major capabilities used to evaluate modern LLMs.

1. What Are LLM Evaluations?

Core Definition

LLM evaluations are systematic, repeatable tests used to judge an LLM or an LLM-powered system against clear evaluation criteria.

LLM evaluations are not simply a collection of random questions asked to a model. They are structured testing systems designed to determine whether a model or LLM-powered application behaves as expected.

The definition contains three critical characteristics:

1. Systematic

Testing follows a structured methodology rather than relying on a few manually selected prompts.

2. Repeatable

The same evaluation can be executed again against different versions of the system.

3. Clear Criteria

The system is evaluated against explicitly defined criteria rather than personal intuition.

2. The Three Core Properties of LLM Evals

Systematic Testing

Systematic testing means that evaluation is performed using a properly prepared dataset rather than casually asking a few questions.

For example, imagine building a chatbot for an educational platform. Instead of manually asking five questions, you could collect a representative sample of real user conversations and convert them into an evaluation dataset.

Goal: Cover normal cases, difficult cases, edge cases, and real-world user behavior as comprehensively as possible.

Repeatable Testing

Repeatability means that the same evaluation dataset should continue to work even when the system changes.

For example, you might change:

  • Prompt
  • LLM
  • Retriever
  • Chunking strategy
  • Embedding model
  • Application logic

Even after these changes, you should be able to run the same evaluation suite and compare the results.

Evaluation Dataset
Version 1
Score
Version 2
Compare

This allows developers to determine whether a new version actually improved the system or introduced regressions.

Clear Evaluation Criteria

The third requirement is defining what "good" actually means for your application.

For example, a chatbot might need to satisfy the following criteria:

  • Answer must be correct.
  • Explanation should be simple and understandable.
  • Answer should use the approved knowledge base.
  • Response should be grounded in retrieved information.
  • Response should be safe.
  • Response should not contain abusive language.
  • Response should maintain the expected tone.

Key idea: Without explicit criteria, you are essentially doing informal testing. With explicit criteria, you are performing structured evaluation.

3. LLM Evals Are Not Just Metrics

A common misunderstanding comes from traditional machine learning.

In classical ML, evaluation often means calculating metrics such as:

  • Accuracy
  • Precision
  • Recall
  • F1 Score
  • ROC-AUC

This can create the misconception that an LLM evaluation is simply a set of metrics.

Important Distinction

An LLM eval is the complete testing setup, not merely a metric.

The complete evaluation setup can include:

  • What component is being tested?
  • What dataset is being used?
  • What criteria are being evaluated?
  • What evaluation method is being used?
  • Which evaluation tools are being used?
  • When is the evaluation performed?
  • Is evaluation offline or in production?
  • How are the results interpreted?

For example, if you are evaluating the retriever inside a RAG system, the retriever, evaluation dataset, retrieval metrics, evaluation framework, and execution process can all be considered parts of the evaluation setup.

4. What Is the Purpose of an LLM Eval?

The goal of an LLM evaluation is not simply to produce a number.

The goal is to answer practical engineering and product questions.

Can We Use This Model?

Is this model capable enough for the task we want to solve?

Is It Ready to Ship?

Is the application reliable enough to be deployed to real users?

Did the Prompt Improve?

Did Prompt Version 2 actually perform better than Prompt Version 1?

Is the RAG Grounded?

Are responses actually supported by the retrieved context?

Is the Agent Correct?

Is the agent completing the intended task correctly?

Is the System Safe?

Can real users safely interact with the system?

Is Latency Acceptable?

Does the application respond quickly enough for the intended user experience?

5. Two Major Types of LLM Evals

LLM evaluations can broadly be divided into two categories:

Type What Is Evaluated? Main Purpose
Model Evals The LLM itself Measure the capabilities and performance of the underlying model.
Application Evals An LLM-powered application or its components Determine whether the complete product or individual components work correctly.

Important terminology note: "Model Evals" and "Application Evals" are useful categories for understanding the topic, but they are not necessarily official industry terms. In practice, both are often discussed broadly as "LLM evaluations," with the context determining which type is meant.

6. Model Evaluations

Model evaluations focus directly on the underlying LLM.

Model Eval

Evaluating the capabilities of an LLM itself using standardized tests and benchmarks.

Whenever a new LLM is released, its creators typically evaluate it across different benchmarks and publish the results.

This helps the industry understand how capable the model is compared with other models.

New LLM
Benchmarks
Evaluation
Scores
Capability Comparison

These evaluations help developers and organizations understand which models might be suitable for their particular applications.

7. Eight Major LLM Capability Categories

Modern LLMs are commonly evaluated across several major capability areas.

Capability 01

Reasoning

Can the model reason through a problem step by step and arrive at an appropriate solution?

Capability 02

Knowledge

Does the model possess sufficient general and world knowledge up to its knowledge cutoff?

Capability 03

Basic Mathematics

Can the model correctly solve mathematical problems?

Capability 04

Coding

Can the model understand programming problems, generate code, and solve coding tasks?

Capability 05

Instruction Following

Can the model follow multiple instructions accurately and consistently?

Capability 06

Long-Context Handling

Can the model retrieve and reason over relevant information from very large contexts?

Capability 07

Multimodal Understanding

Can the model understand and work with different modalities such as text, images, audio, and other forms of input or output?

Capability 08

Tool Use

Can the model correctly use external tools, APIs, functions, or other capabilities?

8. Benchmarks Used for Model Evaluation

Model capabilities are evaluated using standardized benchmarks.

Capability / Area Example Benchmark Purpose
Knowledge & Reasoning MMLU Tests knowledge and reasoning across multiple subjects.
Mathematics GSM8K Tests mathematical problem-solving, especially grade-school-level problems.
Coding SWE-bench Evaluates models on real-world software engineering tasks.
Human Evaluation HumanEval Evaluates code-generation capabilities.
Instruction Following IFEval Tests whether models follow explicit instructions and constraints.
Long Context Needle in a Haystack Tests whether a model can retrieve specific information from a long context.
Multimodal Understanding TRIPLE-MU Used to evaluate multimodal capabilities.

Important: These benchmarks are examples. The benchmark landscape is much larger, and different benchmarks measure different capabilities and behaviors.

9. How Important Are Model Evals for an AI Engineer?

AI engineers should understand model evaluations, but they generally will not be responsible for creating the major industry-wide benchmarks used to evaluate frontier models.

Frontier AI labs typically perform large-scale model evaluations when they release new models.

However, AI engineers should understand:

  • What model evaluations are.
  • What benchmarks are.
  • What different benchmarks measure.
  • How to interpret benchmark results.
  • How benchmark results can influence model selection.

Practical benefit: When starting a project, understanding benchmark results helps you make better decisions about whether to use a proprietary model, another provider's model, or an open-source model.

10. Application Evaluations

Application evaluations are the most important part of this learning path for an AI engineer.

Application Eval

Application evaluations assess the behavior and performance of an LLM-powered application, either at the level of the complete system or at the level of specific components.

A common beginner misconception is:

"The LLM is the brain, so if the LLM is good, the application must also be good."

This is incorrect.

An LLM-powered application usually contains many additional components.

User Interface

The interface through which users interact with the application.

System Prompt

Instructions controlling the model's behavior.

Tools & APIs

External functions, APIs, databases, and services used by the application.

Orchestration

Logic controlling workflows, branching, parallel execution, and agent behavior.

Guardrails

Safety and behavioral controls around the application.

Output Parsers

Components responsible for converting model output into structured formats.

Memory & Context

Mechanisms that maintain relevant information across interactions.

RAG Components

Retrieval systems, embedding models, vector databases, rerankers, and related components.

Monitoring

Production monitoring and observability systems.

Feedback Loop

Mechanisms for collecting user feedback and improving the system.

11. Smartphone Analogy

A useful way to understand application evaluations is through a smartphone analogy.

A smartphone may contain a powerful processor such as a Snapdragon or MediaTek chip.

The processor itself can be benchmarked to determine how powerful it is.

But does having a powerful processor automatically make the smartphone a great product?

Obviously not.

A smartphone also depends on:

  • Camera system
  • Operating system
  • Battery
  • Display
  • Audio system
  • Graphics capabilities
  • Software optimization
  • Connectivity

A powerful processor tells us something about the processor. It does not guarantee that the complete smartphone is a good product.

The same principle applies to LLM applications.

Model Evaluation
+
Application Components
+
Application Evaluation
Reliable Product

Frontier labs evaluate the underlying models. AI engineers are responsible for evaluating the application built around those models.

12. Application Evals: System-Level vs Component-Level

Application evaluations can happen at two levels:

Level What Is Evaluated? Example
System Level The complete LLM-powered application. Is the final RAG chatbot response correct, grounded, safe, fast, and useful?
Component Level Individual components within the application. Is the retriever working correctly? Is the embedding model effective? Is the reranker performing correctly?

For example, in a RAG chatbot you might separately evaluate:

  • Retriever quality
  • Embedding quality
  • Reranker performance
  • Context relevance
  • Answer correctness
  • Groundedness
  • Latency
  • Cost
  • Safety

13. Model Eval vs Application Eval

Question Model Eval Application Eval
What is being evaluated? The underlying LLM. The complete application or one of its components.
Main question "Can the model do this?" "Will our product work correctly?"
Typical owner Frontier model labs. AI/application engineering teams.
Typical tools Standardized benchmarks and leaderboards. Custom datasets, rubrics, evaluators, testing frameworks, and application-specific metrics.
Example Can the LLM solve mathematical problems? Does our RAG chatbot provide grounded and correct answers?

14. Questions Application Evals Should Answer

For an AI engineer, application evaluation should help answer questions such as:

Correctness

Did the application provide the correct answer?

Groundedness

Is the answer supported by the retrieved context?

Completeness

Did the system provide all important information?

Clarity

Is the answer understandable to the intended user?

Safety

Is the application safe for real-world users?

Latency

Does the application respond quickly enough?

Cost

Is the application economically viable at scale?

Task Completion

Can the agent or application complete its intended task correctly?

15. Why Application Evals Matter More to AI Engineers

The job of an AI engineer is to build LLM-powered applications.

Therefore, understanding whether the complete application actually works is a core engineering responsibility.

A model can have excellent benchmark scores while the application built around it can still fail because of:

  • Bad prompts
  • Poor retrieval
  • Incorrect chunking
  • Bad embeddings
  • Incorrect orchestration
  • Tool failures
  • Unsafe outputs
  • Hallucinations
  • High latency
  • High operational cost

A good model does not automatically create a good application.

16. Complete Lecture Summary

This lecture answered the "WHAT" behind LLM evaluations.

Why?

We need LLM evaluations because LLM-powered applications are probabilistic and complex, making traditional software testing insufficient.

What?

LLM evaluations are systematic, repeatable tests performed against clearly defined criteria.

Two Types

Model evaluations measure the capabilities of the underlying LLM, while application evaluations measure the behavior of the application built around it.

How?

The next stage is learning how application evaluations are actually designed, implemented, and executed.

17. Key Takeaways

  1. LLM evals are not just metrics. They represent the complete testing setup.
  2. Good evaluations are systematic. They use structured datasets rather than random manual testing.
  3. Good evaluations are repeatable. The same evaluation should work across different versions of the system.
  4. Evaluation requires clear criteria. You must define what "good" means before testing.
  5. There are two useful categories: model evaluations and application evaluations.
  6. Model evaluations measure model capabilities.
  7. Application evaluations measure the product.
  8. Model evaluation commonly uses benchmarks.
  9. Modern LLMs are evaluated across multiple capabilities including reasoning, knowledge, mathematics, coding, instruction following, long context, multimodality, and tool use.
  10. AI engineers should understand model benchmarks even if they rarely create frontier-model evaluations themselves.
  11. Application evaluation is the core practical focus for AI engineers.
  12. An LLM is only one component of an LLM-powered application.
  13. Application evaluation can happen at both system and component levels.
  14. The key question changes: Model eval asks "Can the model do this?" Application eval asks "Will our product work correctly?"

18. What Comes Next?

Next Topic

How are LLM evaluations actually performed?

The next part moves from the "WHAT" of LLM evaluations toward the "HOW", with the main focus on Application Evaluations.

Why?
What?
Model Evals
Application Evals
How?