Rajan2026/

soas-english-uzbek-rag-evaluation

Bilingual English-Uzbek retrieval-evaluation benchmark for culturally grounded RAG systems. 400 rows (200 EN + 200 UZ), balanced across governance, history, institutions, and culture. Retrieval-only: contains questions and source-document IDs, and excludes answer text. Headline result: targeted Uzbek source curation improved retrieval recall from 39% to 98% (Cohen's d = 2.91).

from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/Rajan2026/soas-english-uzbek-rag-evaluation.