Instruction-Guided Adaptive RAG Evaluation Benchmark

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.