Code and Data: When Retrieval Helps and When It Hurts — A Pre-Registered Cross-Lingual RAG Benchmark for Small Language Models in Greek and Georgian

Replication package for the paper "When Retrieval Helps and When It Hurts: A Pre-Registered Cross-Lingual RAG Benchmark for Small Language Models in Greek and Georgian" (submitted to Baltic Journal of Modern Computing, 2026). Contains: all pipeline scripts (01–11), the final question dataset (238 K-12 science questions in Greek and Georgian), raw model outputs (17,136 generation events across 8 models × 238 questions × 3 conditions × 3 seeds), and judge scores (17,131 doubly-scored answer pairs). The study was pre-registered on Zenodo prior to all benchmark runs: https://doi.org/10.5281/zenodo.20310027 PDF textbooks are not included as they contain copyrighted curriculum material.