LLM Evaluation
Module 9 / 17
09 / 17
Evaluation Methods

LLM Evaluation Methods – Complete Notes (Interview + Production Ready)

This section explains who performs the evaluation in an LLM evaluation pipeline.

Question: Who decides whether the LLM output is good or bad?

The answer is called the Evaluation Method.

What is an LLM Evaluation Method?

Definition

An LLM Evaluation Method is the mechanism used to determine whether an LLM's output is good or bad.

It is the procedure that takes the model output and produces a judgment.

Every evaluation pipeline eventually uses one of three methods.

LLM Output
      │
      ▼
Evaluation Method
      │
      ▼
Score / Pass / Fail

Three Evaluation Methods

                    Evaluation Methods

              ┌───────────────────────────┐
              │ Who performs evaluation?  │
              └─────────────┬─────────────┘
                            │
      ┌─────────────────────┼──────────────────────┐
      │                     │                      │
      ▼                     ▼                      ▼

Programmatic          Human Evaluation       LLM Evaluation
(Deterministic)       (Human Judge)          (LLM-as-a-Judge)

1. Programmatic (Deterministic) Evaluation

Definition

Evaluation is performed by a program.

Python code calculates the metric automatically.

No human judgment is needed during evaluation.

Flow

Dataset
   │
   ▼
LLM Component
   │
   ▼
Python Program
   │
   ▼
Metric

Example

Evaluate the Retriever of a RAG chatbot.

Question:

"What are the prerequisites for the ML course?"

Suppose the correct documents are

1001
1003

Retriever returns Top-5

1001
1002
1004
1005
1006

Recall@K

Formula

Recall@K

=
Relevant Retrieved

----------------------
Total Relevant Documents

Here

Relevant documents

1001
1003

Retrieved relevant

1001

Therefore

Recall@5

=

1
--
2

= 0.5

=50%

Complete Pipeline

Questions
      │
      ▼
Retriever
      │
      ▼
Top-K Documents
      │
      ▼
Python compares
with Ground Truth
      │
      ▼
Recall@K

Advantages

Disadvantages

Cannot evaluate subjective qualities like

Recall@K

Definition

Out of all relevant documents,

how many were retrieved?

Formula

Recall@K

=

Retrieved Relevant Docs

---------------------------

All Relevant Docs

Range

0
↓

1

Higher is better.

Precision@K

Definition

Out of retrieved documents,

how many are actually relevant?

Formula

Precision@K

=

Relevant Retrieved

-----------------------

Retrieved Documents

When to Use Programmatic Evaluation

Use when evaluation is objective.

Examples

2. Human Evaluation

Here,

a human evaluates the output.

Pipeline

Questions
      │
      ▼
Chatbot
      │
      ▼
Generated Answer
      │
      ▼
Human Judge
      │
      ▼
Score

Example

CampusX Chatbot

Question

Can I pay the fee in installments?

Chatbot Answer

↓

Human checks

↓

Assigns score

1–5

Rubric

Example

Score Meaning
5 Accurate, complete, helpful
4 Minor mistakes
3 Partially helpful
2 Poor
1 Completely wrong

Complete Flow

Dataset
      │
      ▼
Chatbot
      │
      ▼
Answers
      │
      ▼
Human reads
      │
      ▼
Assign Score
      │
      ▼
Average Score

Why Multiple Human Evaluators?

Example

Human A → 2

Human B → 5

Large disagreement means

Rubric is ambiguous

Agreement means

Rubric is good.

Human Evaluation Types

1. Direct Rating

Read output

↓

Give score

2. Red Teaming

Humans intentionally try to break the system.

Examples

Purpose

Find vulnerabilities before deployment.

3. A/B Testing

Two chatbot versions

Users

↓

Version A

Version B

↓

Compare ratings

Deploy the better version.

4. Golden Dataset Creation

Humans create

5. Human-in-the-Loop

Complex cases are sent to humans.

Example

LLM uncertain

↓

Human reviews

↓

Final decision

Advantages

Disadvantages

3. LLM Evaluation (LLM-as-a-Judge)

Instead of humans,

another LLM evaluates the output.

This is the most popular evaluation method today.

Why?

Suppose

10,000 UPSC students submit subjective answers.

Human evaluation

Very expensive

Programmatic evaluation

Impossible

Solution

LLM evaluates
all answers.

Pipeline

Student Answer
      │
      ▼
Evaluation Prompt
      │
      ▼
Judge LLM
      │
      ▼
Marks + Reason

Example

Question

Discuss ethical governance.

Rubric

Prompt

You are a UPSSC evaluator.

Use this rubric.

Assign marks.

Explain reasoning.

Output

Score : 12/15

Reason:

Good examples.

Weak conclusion.

How to Evaluate the Judge LLM?

Need a Golden Dataset.

Student Answer

↓

Human Marks

↓

LLM Marks

↓

Compare

Example

Answer Human LLM
A1 13 12
A2 8 9
A3 4 4

Metric

Mean Absolute Error (MAE)

Formula

MAE

=

Σ |Human Score - LLM Score|

-----------------------------

Number of Answers

Example

Human

13
8
4

LLM

12
9
4

Difference

1

1

0

MAE

(1+1+0)/3

=0.67

Lower MAE is better.

MAE = 0

Perfect Judge

Advantages

Disadvantages

Comparison

Feature Programmatic Human LLM Judge
Cost Very Low High Medium
Speed Very Fast Slow Fast
Scalable Yes No Yes
Subjective Tasks No Yes Yes
Reliability High (objective) Highest Good

Which Method Should You Use?

Objective Metric?
        │
        ▼
Programmatic

Need Human Judgment?
        │
        ▼
Can afford Humans?

        │
   Yes ─────► Human

        │
       No
        ▼
LLM-as-a-Judge

Reference-Based Evaluation

Definition

Correct answer already exists.

Model output is compared against the reference.

Example

Retriever Evaluation

Question

↓

Correct document IDs already known

↓

Compute Recall

Another Example

Human marks

13

LLM gives

12

Compare both.

Reference already exists.

Reference-Free Evaluation

No correct answer exists.

Judge evaluates quality directly using a rubric.

Example

Question

↓

Chatbot Answer

↓

Human reads

↓

Rate Helpfulness

No predefined correct answer.

Difference

Reference-Based Reference-Free
Ground truth available No ground truth
Compare against reference Judge directly
Objective Subjective
Uses labels Uses rubric

Complete Evaluation Pipeline

Step 1

Define Task

↓

Step 2

Define Success Metric

↓

Step 3

Create Evaluation Dataset

↓

Step 4

Choose Evaluation Method

Program
Human
LLM

↓

Step 5

Run Evaluation

↓

Step 6

Calculate Metrics

↓

Step 7

Analyze Failures

↓

Step 8

Improve System

↓

Repeat

Interview Questions

Q1. What are the three LLM evaluation methods?

Q2. When should Programmatic Evaluation be used?

When outputs can be measured objectively (accuracy, Recall@K, Exact Match, JSON validity, etc.).

Q3. Why is Human Evaluation considered the gold standard?

Because humans best understand context, nuance, reasoning, tone, and helpfulness.

Q4. Why use LLM-as-a-Judge?

It offers a balance between scalability and quality—far cheaper than humans while handling subjective evaluation better than deterministic metrics.

Q5. What is Recall@K?

The fraction of all relevant documents that appear within the top K retrieved results.

Q6. What is MAE in LLM evaluation?

Mean Absolute Error measures the average absolute difference between human scores and LLM-generated scores. Lower MAE indicates closer agreement.

Q7. What is a Golden Dataset?

A manually curated dataset containing trusted ground-truth labels, answers, or scores used as the benchmark for evaluation.

Q8. What is a Rubric?

A predefined set of evaluation criteria (e.g., accuracy, completeness, reasoning, tone) used to consistently score outputs.

Q9. What is Reference-Based Evaluation?

Evaluation where a known correct answer or label exists and model output is compared against it.

Q10. What is Reference-Free Evaluation?

Evaluation where no ground-truth answer exists; quality is judged directly using criteria or a rubric.

Absolutely—I've converted this lecture into structured, interview-ready notes with the same teaching flow, diagrams, examples, formulas, and important points.

You can use these notes directly for LLM Evaluation, RAG Evaluation, LLM-as-a-Judge, and AI Engineer interviews.

LLM Evaluation Methods (Complete Notes)

What is an Evaluation Method?

Definition

An LLM Evaluation Method is the mechanism that decides whether an LLM's output is good or bad.

It is simply who performs the evaluation.

LLM Output
      │
      ▼
Evaluation Method
      │
      ▼
Good / Bad
Score / Pass / Fail

Every LLM evaluation pipeline eventually uses only one of these three methods.

Three Evaluation Methods

                 LLM Evaluation

                     │
      ┌──────────────┼──────────────┐
      │              │              │
      ▼              ▼              ▼

Programmatic      Human         LLM-as-a-Judge
(Deterministic) Evaluation      (Model-Graded)

1. Programmatic (Deterministic) Evaluation

Definition

A program (Python code) performs the evaluation.

No human judgment is involved during evaluation.

The evaluation is completely deterministic.

When is it used?

Whenever quality can be measured mathematically.

Examples

Example

Evaluate a RAG Retriever

Retriever's job

Question
    │
    ▼
Retriever
    │
    ▼
Relevant Documents

Question

What are the prerequisites for ML Course?

Ground Truth

1001
1003

Retriever returns Top-5

1001
1002
1004
1005
1006

Recall@K

Definition

Out of all relevant documents,

how many were retrieved?

Formula

Recall@K

=

Relevant Retrieved

------------------------
Total Relevant Documents

Example

Relevant documents

1001
1003

Retrieved relevant

1001

Therefore

Recall@5

=

1
──
2

=

0.5

=

50%

Evaluation Pipeline

Evaluation Dataset
      │
      ▼
Retriever
      │
      ▼
Top-K Documents
      │
      ▼
Python compares
with Ground Truth
      │
      ▼
Recall@K

Complete Pipeline

Step 1

Define Task

Evaluate Retriever

Step 2

Success Metric

Recall@K

Step 3

Create Evaluation Dataset

Example

Question Correct Documents
Q1 1001
Q2 1001,1003
Q3 1005

This is called the Golden Dataset.

Step 4

Run Retriever

Question

↓

Retriever

↓

Top-K Docs

Step 5

Program Calculates Recall

Example

Q1

Correct

1001

Retrieved

1001

Recall

1

Q2

Correct

1001
1003

Retrieved

1001

Recall

0.5

Average Recall

Recall@5

=

Average of all recalls

How to Improve Recall?

Advantages

✅ Fast

✅ Cheap

✅ Repeatable

✅ Fully Automated

Disadvantages

Cannot evaluate

2. Human Evaluation

Instead of a program,

a human performs the evaluation.

Pipeline

Question

↓

Chatbot

↓

Answer

↓

Human Judge

↓

Score

Example

Question

Can I pay fee in installments?

Chatbot generates answer.

Human checks

Then gives score.

Rubric

Example

Score Meaning
5 Excellent
4 Good
3 Average
2 Poor
1 Wrong

Human Evaluation Flow

Evaluation Dataset

↓

Questions

↓

Chatbot

↓

Generated Answers

↓

Human reads

↓

Assign Score

↓

Average Score

Why Multiple Human Evaluators?

Suppose

Evaluator A

2

Evaluator B

5

Large disagreement means

Rubric is ambiguous.

If both give similar scores

Rubric is reliable.

This is called Inter-Annotator Agreement.

Types of Human Evaluation

1. Direct Rating

Human directly gives score.

Read

↓

Judge

↓

Score

2. Red Teaming

Humans intentionally attack the LLM.

Goal

Find

3. A/B Testing

Two chatbot versions

Users

↓

Version A

Version B

↓

Ratings

↓

Choose Better Version

4. Golden Dataset Creation

Humans create

5. Human-in-the-Loop

Complex cases

↓

Sent to Human

↓

Human decides final output

Advantages

Highest reliability

Best reasoning

Handles ambiguity

Disadvantages

Very expensive

Slow

Not scalable

Needs experts

3. LLM-as-a-Judge (Model-Graded Evaluation)

Instead of humans,

another LLM evaluates outputs.

This is currently the most popular evaluation method in production LLM systems.

Why?

Suppose

10,000 UPSC students submit subjective answers.

Human evaluation

❌ Very expensive

Programmatic evaluation

❌ Impossible

Solution

Student Answer

↓

Judge LLM

↓

Marks

Example

Question

Discuss Ethical Governance

Rubric

Prompt

You are a UPSC evaluator.

Use the following rubric.

Assign marks.

Explain your reasoning.

Output

Marks

12/15

Reason

Good explanation

Weak conclusion

Evaluation Pipeline

Student Answer

↓

Judge Prompt

↓

Judge LLM

↓

Marks

+

Reasoning

How to Evaluate the Judge LLM?

Need a Golden Dataset.

Example

Answer Human Marks LLM Marks
A1 13 12
A2 4 5
A3 8 8

Now compare both.

Metric

Mean Absolute Error (MAE)

Formula

MAE

=

Σ |Human − LLM|

-------------------

Number of Answers

Example

Human

13
8
4

LLM

12
9
4

Difference

1

1

0

MAE

(1+1+0)

────────

3

=

0.67

Lower is better.

Ideal

MAE

=

0

How to Improve Judge LLM?

Advantages

Much cheaper than humans

Fast

Scalable

Reasoning capable

Disadvantages

Can hallucinate

May inherit bias

Prompt sensitive

Less reliable than experts

Comparison

Feature Programmatic Human LLM Judge
Speed ⭐⭐⭐⭐⭐ ⭐ ⭐⭐⭐⭐
Cost Very Low Very High Medium
Scalability Excellent Poor Excellent
Subjective Tasks No Yes Yes
Reliability High (objective) Highest Good

Which Evaluation Method Should You Use?

Objective Metric?

↓

Yes

↓

Programmatic



Need Human Judgment?

↓

Can afford Humans?

↓

Yes

↓

Human



No

↓

LLM-as-a-Judge

Reference-Based Evaluation

Definition

A correct answer already exists.

Evaluation compares model output with this reference.

Examples

Ground Truth

↓

Model Output

↓

Comparison

↓

Score

Reference-Free Evaluation

Definition

No predefined correct answer exists.

The evaluator judges quality directly using a rubric.

Question

↓

Chatbot Answer

↓

Judge

↓

Helpfulness

Examples

Reference-Based vs Reference-Free

Reference-Based Reference-Free
Ground truth available No ground truth
Compare with reference Judge directly
Objective Subjective
Uses labels Uses rubrics

Complete LLM Evaluation Pipeline

Step 1

Define Task

↓

Step 2

Choose Target

(Component / Workflow / Application)

↓

Step 3

Define Success Metric

↓

Step 4

Create Evaluation Dataset

↓

Step 5

Choose Evaluation Method

Programmatic

Human

LLM Judge

↓

Step 6

Run Evaluation

↓

Step 7

Compute Metrics

↓

Step 8

Analyze Failures

↓

Step 9

Improve System

↓

Repeat

Interview Questions

1. What are the three LLM evaluation methods?

2. When should Programmatic Evaluation be used?

When output quality can be measured objectively using metrics such as Accuracy, Recall@K, Precision, Exact Match, or JSON validation.

3. Why is Human Evaluation considered the gold standard?

Because humans best understand context, nuance, reasoning, tone, helpfulness, and subjective quality.

4. Why is LLM-as-a-Judge widely used?

It provides a practical balance between cost, speed, scalability, and evaluation quality, making it suitable for production systems.

5. What is Recall@K?

The fraction of all relevant documents that appear in the top K retrieved results.

6. What is MAE?

Mean Absolute Error measures the average absolute difference between human-assigned scores and LLM-assigned scores. A lower MAE indicates better agreement.

7. What is a Golden Dataset?

A curated dataset containing trusted ground-truth labels, answers, or scores used as the benchmark for evaluation.

8. What is a Rubric?

A predefined set of scoring criteria (e.g., accuracy, completeness, reasoning, examples, tone) used to evaluate outputs consistently.

9. What is Reference-Based Evaluation?

Evaluation where a known correct answer or label exists and model output is compared against that reference.

10. What is Reference-Free Evaluation?

Evaluation where no ground-truth answer exists; outputs are judged directly against a rubric or quality criteria.

These notes capture the complete lecture in a structured, production-oriented format while preserving the concepts, examples, formulas, diagrams, and interview-relevant details.

I'm glad the notes helped! This lecture covers one of the most important topics in LLM Evaluation. The notes above already include the complete explanation in a structured format, so you can use them directly for revision and interviews.

The next topic in this course, as mentioned at the end of the lecture, is:

Offline vs Online Evaluation

This is another core production concept. If you share that lecture transcript as well, I'll convert it into notes in the same style:

That way, you'll have a complete LLM Evaluation Handbook covering every topic in a consistent format.

Perfect! Send the next lecture, and I'll convert it into the same detailed, production-ready notes with:

I'll maintain the same format throughout so that, by the end, you'll have a complete LLM Evaluation Notes / Handbook suitable for interviews and revision.