gaffarshaikh07/rag-qa-evaluation-dataset

---

pretty_name: RAG QA Evaluation Dataset language: - en license: mit tags: - llm-evaluation - rag - ai-testing - qa - hallucination - question-answering task_categories: - question-answering size_categories: - n<1K

RAG QA Evaluation Dataset

Overview

This dataset contains test cases for evaluating Retrieval-Augmented Generation (RAG) and Large Language Model (LLM) applications.

The dataset is designed from a software testing and quality engineering perspective.

Dataset Structure

Each test case contains:

Field	Description
question	User question sent to the AI application
context	Context available to the AI application
expected_answer	Expected behavior or answer
test_type	Category of the test case

Test Types

Grounded

Questions that can be answered directly from the supplied context.

Boundary

Questions that test limits and edge cases.

Out-of-context

Questions where the answer cannot be determined from the supplied context.

The expected behavior is for the application to avoid inventing unsupported information.

Prompt Injection

Tests designed to check whether an application follows its intended instructions and context when presented with unrelated instructions.

Intended Use

This dataset can be used for:

• RAG testing
• LLM evaluation
• AI application testing
• QA automation
• Regression testing
• Hallucination testing
• Negative testing
• Boundary testing

Potential evaluation metrics include:

• Faithfulness
• Answer Relevancy
• Contextual Relevancy
• Hallucination
• Task Completion
• Tool Correctness
• Argument Correctness

Example Test Case

Question:

What is the maximum age to purchase the policy?

Context:

The policy is available to customers aged 18 to 65 years.

Expected answer:

65 years

Test type:

grounded

Limitations

This is a small demonstration and evaluation dataset. It is not intended to represent a comprehensive benchmark for production RAG or LLM systems.

The examples are designed primarily for learning, experimentation, and software testing demonstrations.

Data Quality

The test cases were manually designed to represent common AI application testing scenarios including positive, boundary, negative, and out-of-context cases.

License

MIT