StructFlowBench
收藏资源简介:
StructFlowBench是由吉林大学提出的一个针对多轮指令遵循能力的结构化流评估基准。该数据集包含8种任务类型、22个主题和13种约束类型,共计155个多轮对话,643个话轮和1,775个约束。它通过结构化流框架,将六种基本的话轮关系(追随、细化、回忆、总结、扩展、无关)融入多轮对话,旨在为多轮对话的指令遵循评估提供一个可解释的框架,并促进对话系统的结构化理解和分析。
StructFlowBench is a structured flow evaluation benchmark proposed by Jilin University for evaluating multi-turn instruction-following performance. This dataset covers 8 task categories, 22 topics, and 13 constraint types, with a total of 155 multi-turn dialogues, 643 utterances, and 1,775 constraints. It integrates six basic utterance relationship types (follow-up, refinement, recall, summarization, expansion, and irrelevant) into multi-turn dialogues through a structured flow framework, aiming to provide an interpretable framework for instruction-following evaluation in multi-turn dialogues and facilitating structured understanding and analysis of dialogue systems.

- 1StructFlowBench: A Structured Flow Benchmark for Multi-turn Instruction Following吉林大学 · 2025年



