ilsp/Greek_Myths_POIs_RAG

Greek Myths–POIs RAG Dataset

Dataset Description

The Myths–POIs RAG QA dataset is a Greek-language question-answering dataset for evaluating Retrieval-Augmented Generation (RAG) systems in the cultural domain.

The dataset is based on cultural resources curated by the Institute for Language and Speech Processing (ILSP) from the Mythotopia collection and covers two domains:

• Greek Myths
• Greek Points of Interest (POIs)

Each instance consists of a question, a reference answer, and its corresponding source context.

Dataset Creation

QA pairs were synthetically generated from the source texts using GPT-4o mini and Llama-Krikri-8B-Instruct. Generation was restricted to the information available in the provided context.

Approximately 5.200 QA pairs were initially generated. Automated filtering was subsequently applied to remove unsuitable content, generic question formulations, and redundant or highly similar questions, resulting in approximately 940 candidate QA pairs.

Human Annotation and Validation

The candidate QA pairs were evaluated by 3 human annotators using Argilla. Each instance was assessed on a 0–3 ordinal scale across five dimensions:

• Context Relevance
• Context Sufficiency
• Answer Correctness
• Answer Faithfulness
• Answer Completeness

Annotators could also correct questions and answers when necessary. Following the annotation and quality assessment process, 670 QA pairs were retained as validated instances, of which 506 are included in the publicly released dataset.

As an additional validation step, the retained instances were evaluated using Qwen/Qwen-2.5-72B-Instruct as an LLM judge. Cases presenting potential inconsistencies or quality issues were manually reassessed.

Human–LLM Agreement

Metric	Human Mean	LLM Mean	Exact Agreement	MAE
Context Relevance	2.9955	3.0000	99.55%	0.0045
Context Sufficiency	2.9731	3.0000	97.61%	0.0269
Answer Correctness	2.9567	2.9955	95.37%	0.0478
Answer Faithfulness	2.9164	2.9940	91.64%	0.0896
Answer Completeness	2.8612	2.9627	85.97%	0.1463

Overall, exact human–LLM agreement ranged from 85.97% to 99.55%, with the highest agreement observed for context relevance and the largest differences for answer completeness.

Intended Use

The dataset is intended for the evaluation of Greek-language RAG and question-answering systems.

Citation

If you use this dataset in your research or project, please consider including the following citation:

@misc{greek_myths_pois_rag_qa,
  title={Greek Myths--POIs RAG QA},
  author={Batsi, Eleni},
  url={https://huggingface.co/datasets/ilsp/Greek_Myths_POIs_RAG_QA},
  year={2026}
}