2026-10-15T01:02:18Z https://www.scidb.cn/oai

10.57760/sciencedb.00zzs 2026-09-20T09:57:25Z

2026-09-20 Instruction-Guided Adaptive RAG Evaluation Benchmark doi:10.57760/sciencedb.00zzs en This dataset records the experimental dataset corresponding to the paper "IG-RAG: Retrieval Enhancement Generation Method Integrating Difficulty Aware Retrieval and Instruction Guided Evidence Utilization". Including the main result data of the paper experiment; Six publicly available datasets used: SQuAD, Natural Questions, TriviaQA, MuSiQue, HotpotQA, 2WikiMultiHopQA; Prompt words corresponding to the dataset; A dataset of 60000 self synthesized inference trajectories. And representative complex query examples were carefully selected from general and domain Q&A scenarios, and structured queries were expanded by combining different task instruction constraints. High precision retrievers were used to match multi granularity related document paragraphs, and finally, a processed dataset containing 500 entries was generated through quality filtering and cross validation, ensuring the authority and high accuracy of each data entry.  LLMs; RAG; Adaptive Retrieval; Difficulty‑aware Splitting; Evaluation Benchmark Wang Hao Chen Mingkun Zhao Liang Li Jin PUBLIC https://creativecommons.org/licenses/by-nc/4.0/ dataset Science Data Bank