---
language:
- en
license:
- mit
multilinguality:
- monolingual
size_categories:
- n<1K
tags:
- rag
- retrieval-augmented-generation
- evaluation
- benchmark
- question-answering
- faithfulness
pretty_name: RAG Evaluation Benchmark & QA Test Suite
---

# RAG Evaluation Benchmark & QA Test Suite 📚

A curated ground-truth evaluation benchmark designed to assess Retrieval-Augmented Generation (RAG) pipelines across three primary dimensions:
1. **Context Recall**: Did the retriever fetch the passage containing the necessary ground-truth facts?
2. **Faithfulness**: Is the generated answer supported strictly by the retrieved evidence without hallucinated extrapolations?
3. **Answer Relevancy**: Does the generated synthesis directly answer the question?

- **Associated Space**: [RAG Engine Architecture Showcase](https://huggingface.co/spaces/Jayan12321/rag-engine)
- **GitHub Repository**: [JayanGupta/RAG-Engine](https://github.com/JayanGupta/RAG-Engine)
- **Author**: [Jayan Gupta](https://huggingface.co/Jayan12321)

---

## 📊 Dataset Schema

| Field | Type | Description |
|---|---|---|
| `question_id` | string | Unique identifier for the evaluation item. |
| `question` | string | User query spanning technical domains (quantum, space, clean energy). |
| `ground_truth_answer` | string | Verified, objective factual answer derived from the source knowledge corpus. |
| `evidence_doc` | string | Title of the source document containing the authoritative facts. |
| `question_type` | string | Type of cognitive task: `factoid`, `synthesis`, or `comparative`. |

---

## 💻 Quickstart with Python

```python
from datasets import load_dataset

# Load directly from Hugging Face
dataset = load_dataset("Jayan12321/rag-evaluation-benchmark")

# Inspect sample evaluation question
sample = dataset["train"][0]
print("Question:", sample["question"])
print("Ground Truth:", sample["ground_truth_answer"])
print("Evidence Document:", sample["evidence_doc"])
```

---

## 🛠️ Usage in RAG Benchmarking
This dataset is designed to be paired with heuristic faithfulness checkers or RAGAS-style evaluation frameworks to objectively score retriever and generator configurations.