Xavier-Xia/sptt-benchmark-data

SPTT基准数据集是为论文《Streaming Propagation Through Time for Training Recurrent Neural Networks》生成的，旨在支持长序列推理和控制性体制切换压力测试实验的可复现性。数据集包含两个主要部分：1. long_listops/：包含生成的Long ListOps数据，用于长序列序列分类实验。ListOps是一个合成序列推理任务，最初用于评估长程依赖建模，该数据以Parquet格式存储，以减少存储大小并提高加载效率，旨在评估循环训练方法是否能处理长符号序列并在扩展的时间范围内保持有用的学习信号。2. regime_switch/：包含一个合成体制切换时间序列数据集，设计用于在非平稳动态下对低秩循环学习方法进行压力测试，用于评估SPTT在训练期间主导学习子空间变化时是否能适应。数据集包含两个体制，在指定步骤切换以改变基础时间结构，迫使模型在体制转换后适应其学习的循环表示。该数据集适用于分析非平稳时间动态的适应性、体制变化后的子空间漂移、对秩参数的敏感性、低秩循环学习信号的稳定性以及奇异值轨迹和每模式贡献的行为。所有生成的数据集均以Parquet格式存储，提供紧凑的存储和高效的列访问。公共数据集如Time-Series-Library、AG News、MNIST和CIFAR-10未包含在此仓库中，需从原始公共源下载。The SPTT Benchmark Data repository contains generated datasets used in the paper Streaming Propagation Through Time for Training Recurrent Neural Networks to support reproducibility of experiments involving long-sequence reasoning and controlled regime-switch stress tests. The dataset consists of two main parts: 1. long_listops/: This directory contains generated Long ListOps data for long-sequence sequence-classification experiments. ListOps is a synthetic sequence reasoning task originally introduced to evaluate long-range dependency modeling. The data are stored in Parquet format to reduce storage size and improve loading efficiency, and are intended for evaluating whether recurrent training methods can handle long symbolic sequences and maintain useful learning signals over extended temporal horizons. 2. regime_switch/: This directory contains a synthetic regime-switch time-series dataset designed for stress-testing low-rank recurrent learning methods under non-stationary dynamics. It was used to evaluate whether SPTT can adapt when the dominant learning subspace changes during training. The dataset includes two regimes that switch at a specified step to change the underlying temporal structure, forcing the model to adapt its learned recurrent representation after the regime transition. It is useful for analyzing adaptation to non-stationary temporal dynamics, subspace drift after a regime change, sensitivity to the rank parameter, stability of low-rank recurrent learning signals, and behavior of singular-value trajectories and per-mode contribution. All generated datasets are stored in Parquet format for compact storage and efficient columnar access. Public datasets such as Time-Series-Library, AG News, MNIST, and CIFAR-10 are not redistributed in this repository and should be downloaded from their original public sources.