PBIF
收藏资源简介:
PBIF数据集是由复旦大学数据科学研究所创建,用于研究多约束指令遵循中的位置偏差问题。该数据集包含24,000个样本,是通过合成多种任务和约束组合生成的多约束指令。数据集的构建考虑了不同类型的约束,并以难度分布指数(CDDI)来量化不同约束顺序的难度差异。该数据集旨在帮助理解和改善大型语言模型在处理多约束指令时的性能表现。
The PBIF dataset was developed by the Institute of Data Science at Fudan University for research on the positional bias problem in multi-constrained instruction following. It contains 24,000 samples, which are multi-constrained instructions generated by synthesizing diverse combinations of tasks and constraints. The dataset construction accounts for various types of constraints, and adopts the Constraint Difficulty Distribution Index (CDDI) to quantify the difficulty differences across different constraint orders. This dataset is designed to facilitate the understanding and performance improvement of large language models (LLMs) when handling multi-constrained instructions.

- 1Order Matters: Investigate the Position Bias in Multi-constraint Instruction Following复旦大学数据科学研究所 · 2025年



