PeatRAG RAGAS Evaluation Dataset: A Held-Out Question-Answer Benchmark for Multilingual Peatland Policy RAG Evaluation

This dataset is the held-out evaluation benchmark for the RAG-based question-answering tool integrated into the European Peatland Policy Portal (https://peatlandpolicyportal.eu), developed under the EU LIFE MultiPeat project (LIFE20 CCM/DE/001802). The dataset contains 58 question–answer pairs generated automatically from 12 peatland-specific policy documents indexed in the portal's Qdrant vector database, using the RAGAS TestsetGenerator framework (ragas v0.1.21). Questions and ground truth answers were generated by Mistral 7B from real policy document chunks; they are grounded in actual indexed content but are model-generated rather than manually annotated by domain experts. COVERAGE- Countries: Belgium, Bulgaria, Germany, Greece, Ireland, Latvia, Romania, Netherlands United Kingdom (9 countries)- Governance levels: Supranational (EU), National, Regional, Local (4 levels)- Languages of source documents: English, German, Flemish- Question types: Targets/Numeric - 24 questions (41%) Dates - 14 questions (24%) Purpose/Definition-13 questions (22%) Location - 7 questions (12%) RAGAS EVALUATION RESULTS - Context Precision: 0.895 (overall); 0.978 (Dates), 0.940 (Targets), 0.851 (Purpose), 0.657 (Location)- Answer Relevancy: 0.638 (overall); 0.642 (Dates), 0.614 (Targets), 0.828 (Purpose), 0.406 (Location)- Citation Precision: 1.000 (overall)- LLM backend: Mistral 7B via Ollama (local inference, no external API)- Evaluation framework: RAGAS v0.1.21 FILE CONTENTS- PeatRAG_RAGAS_Evaluation_Dataset_v4.xlsx - The evaluation dataset (58 Q&A pairs, question type labels, evolution type, source document reference)- PeatRAG_RAGAS_Results_Final.xlsx - Evaluation results with per-row scores, per-type summary table, and evaluation notes- generate_testset.py - Script used to generate the dataset from Qdrant- peatrag_ragas_eval.py - Script used to run RAGAS evaluation GENERATION METHODQuestions were generated using the RAGAS TestsetGenerator with four evolution strategies (simple, conditional, reasoning, multi_context) applied to stratified samples of document chunks drawn from the portal's Qdrant nlp_docs collection, filtered to peatland-specific documents only. A text quality filter (minimum 50% Unicode letter character ratio, minimum 10 words) was applied to exclude garbled PDF extractions before generation. INTENDED USEThis dataset is intended for:- Replication and extension of the RAG evaluation reported in Waskow et al. (2026)- Benchmarking of other RAG systems on multilingual environmental policy QA- Development of evaluation protocols for domain-specific policy QA systems