SalutefaceAAC/rag-evaluation-pairs

RAG Evaluation Pairs

A hand-authored dataset of 122 context–question–answer triples designed for evaluating and training RAG (Retrieval-Augmented Generation) systems. Each example includes richly annotated quality scores covering faithfulness, relevance, and completeness, along with structured error-type labels.

Dataset Summary

RAG evaluation is a critical bottleneck for deploying reliable AI systems: it is easy to generate answers but hard to systematically assess whether they are grounded, complete, and relevant. This dataset provides a labeled evaluation benchmark that covers the full spectrum of answer quality — from perfectly faithful responses to outright hallucinations and irrelevant answers.

Split	Examples
train	97
validation	12
test	13
Total	122

Dataset Fields

Field	Type	Description
context	string	The retrieved passage the answer should be grounded in
question	string	The question posed to the RAG system
candidate_answer	string	The candidate answer to evaluate
faithfulness_score	int (1–5)	How well-grounded the answer is in the context
relevance_score	int (1–5)	How relevant the answer is to the question
completeness_score	int (1–5)	How completely the answer covers the relevant information
verdict	string	correct / partial / incorrect
explanation	string	Human-authored explanation of the scores
error_type	string	Error category (see below)
domain	string	Knowledge domain
difficulty	string	easy / medium / hard

Score Rubric

Faithfulness (1–5)

• 5: Every claim is directly supported by the context
• 4: Mostly supported; minor inferences beyond context
• 3: Mixed — some claims supported, others not
• 2: Mostly unsupported or significantly paraphrased incorrectly
• 1: Contradicts or entirely ignores the context

Relevance (1–5)

• 5: Directly answers the question asked
• 3: Tangentially related but misses the core question
• 1: Completely unrelated to the question

Completeness (1–5)

• 5: Covers all relevant information from the context
• 3: Partially covers the relevant content
• 1: Missing most relevant information

Error Types

Error Type	Description	Count
none	Answer is correct and faithful	83
hallucination	Answer introduces facts not in context	20
contradictory	Answer contradicts the context	12
irrelevant	Answer doesn't address the question	4
vague	Answer is too vague to be useful	3

Domain Distribution

Domain	Count
technology	30
environment	13
history	12
medical	11
finance	9
biology	8
physics	8
legal	7
chemistry	7
arts	7
neuroscience	7
literature	3

Sample Records

Example 1 — Correct (Faithfulness: 5)

{
  "context": "Mitochondria are organelles found in eukaryotic cells. They produce adenosine triphosphate (ATP) through a process called cellular respiration, specifically via oxidative phosphorylation in the inner mitochondrial membrane...",
  "question": "What process do mitochondria use to produce ATP?",
  "candidate_answer": "Mitochondria produce ATP through oxidative phosphorylation, which occurs in the inner mitochondrial membrane during cellular respiration.",
  "faithfulness_score": 5,
  "relevance_score": 5,
  "completeness_score": 5,
  "verdict": "correct",
  "explanation": "The answer is fully supported by the context and directly answers the question with accurate detail.",
  "error_type": "none",
  "domain": "biology",
  "difficulty": "easy"
}

Example 2 — Hallucination (Faithfulness: 1)

{
  "context": "Mitochondria are organelles found in eukaryotic cells... Mitochondria have their own DNA and ribosomes...",
  "question": "How many mitochondria does a typical human cell contain?",
  "candidate_answer": "A typical human cell contains between 1,000 and 2,500 mitochondria, depending on the cell's energy demands.",
  "faithfulness_score": 1,
  "relevance_score": 4,
  "completeness_score": 3,
  "verdict": "incorrect",
  "explanation": "The context does not mention the number of mitochondria per cell. The answer introduces specific numbers (1,000–2,500) not found in the context — this is hallucination.",
  "error_type": "hallucination",
  "domain": "biology",
  "difficulty": "medium"
}

Example 3 — Contradictory (Faithfulness: 1)

{
  "context": "The Treaty of Versailles was signed on June 28, 1919...",
  "question": "When was the Treaty of Versailles signed?",
  "candidate_answer": "The Treaty of Versailles was signed in 1918, shortly after the armistice ended the fighting.",
  "faithfulness_score": 1,
  "relevance_score": 4,
  "completeness_score": 3,
  "verdict": "incorrect",
  "explanation": "The context clearly states the treaty was signed on June 28, 1919. The answer gives the wrong year (1918).",
  "error_type": "contradictory",
  "domain": "history",
  "difficulty": "easy"
}

Use Cases

1. RAG evaluation pipeline: Use as a labeled test set to benchmark your RAG system's faithfulness, relevance, and completeness.
2. Hallucination detection: Train classifiers to detect when LLM answers stray from provided context.
3. Answer quality scoring: Fine-tune reward models or scoring heads on the tri-score annotations.
4. Prompt engineering: Use error-type labels to identify systematic failure modes (hallucination vs. vagueness vs. irrelevance).
5. Teaching/education: Use as examples in courses on RAG system evaluation.

Methodology

Examples were hand-authored with the following design principles:

• Realistic contexts: Passages are factual, multi-sentence paragraphs from diverse knowledge domains — the kind of chunks a retrieval system would return.
• Targeted error injection: Incorrect answers were deliberately constructed to test specific failure modes (hallucination, contradiction, vagueness, irrelevance), not just "bad" answers at random.
• Score granularity: Three independent scores (not one overall score) enable fine-grained analysis of failure modes.
• Difficulty calibration: hard examples often require synthesis across the context, not just extraction, or involve subtle errors that are hard to spot.
• Negative class is genuine: Questions where the context simply does not contain the answer are represented — and the "correct" response is to say so.

Limitations

• 122 examples is a benchmark-scale dataset, not a large training corpus. Use for evaluation, not extensive fine-tuning.
• All contexts are synthetic/curated; domain coverage is non-exhaustive.
• Scores reflect one annotator's judgment; inter-annotator agreement was not measured.
• Contexts are English-only.

Citation

@dataset{salutefaceaac_rag_eval_2026,
  author = {SalutefaceAAC},
  title = {RAG Evaluation Pairs},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/SalutefaceAAC/rag-evaluation-pairs},
  license = {Apache-2.0}
}

License

Apache 2.0 — free to use for research and commercial purposes.