README.md

---
license: cc-by-4.0
task_categories:
  - time-series-forecasting
  - text-classification
  - tabular-regression
language:
  - en
pretty_name: SPTT Benchmark Data
tags:
  - recurrent-neural-networks
  - long-sequence-modeling
  - streaming-learning
  - low-rank-learning
  - listops
  - synthetic-data
  - regime-switch
size_categories:
  - 100K<n<1M
---

# SPTT Benchmark Data

This repository contains generated datasets used in the paper:

**Streaming Propagation Through Time for Training Recurrent Neural Networks**

The datasets are provided to support reproducibility of the experiments involving long-sequence reasoning and controlled regime-switch stress tests. Public datasets used in the paper, such as Time-Series-Library, AG News, MNIST and CIFAR-10, are **not redistributed** in this repository. They should be downloaded from their original public sources using the scripts provided in the accompanying code repository.

## Repository structure

```text
sptt-benchmark-data/
├── long_listops/
│   ├── basic_train.parquet
│   ├── basic_vali.parquet
|   └── basic_test.parquet
├── regime_switch/
│   ├── regime_switch.parquet
│   └── meta.txt
└── README.md
```

## Dataset contents

### 1. `long_listops/`

The `long_listops/` directory contains generated Long ListOps data used for long-sequence sequence-classification experiments.

ListOps is a synthetic sequence reasoning task originally introduced to evaluate long-range dependency modeling. In this repository, the generated Long ListOps files are stored in Parquet format to reduce storage size and improve loading efficiency.

The data 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/`

The `regime_switch/` directory contains a synthetic regime-switch time-series dataset designed for stress-testing low-rank recurrent learning methods under non-stationary dynamics.

This dataset was used to evaluate whether SPTT can adapt when the dominant learning subspace changes during training.

The metadata for the generated dataset are:

```text
total_steps: 640000
switch_step: 320000
feature_dim: 7
noise_std: 0.05
seed: 2025
target_column: f1
```

The sequence contains two regimes. The first regime appears before `switch_step`, and the second regime appears after `switch_step`. The switch is designed to change the underlying temporal structure, so that the model must adapt its learned recurrent representation after the regime transition.

The dataset is useful for analysing:

- adaptation to non-stationary temporal dynamics;
- subspace drift after a regime change;
- sensitivity to the rank parameter;
- stability of low-rank recurrent learning signals;
- behaviour of singular-value trajectories and per-mode contribution.

## File format

All generated datasets are stored in **Parquet** format.

Parquet was chosen because it provides compact storage, efficient columnar access and compatibility with common Python data-processing libraries.

Example loading code:

```python
import pandas as pd

df = pd.read_parquet("regime_switch/regime_switch.parquet")
print(df.head())
```

or, using Hugging Face Datasets:

```python
from datasets import load_dataset

long_listops = load_dataset("Xavier-Xia/sptt-benchmark-data", "long_listops")
regime_switch = load_dataset("Xavier-Xia/sptt-benchmark-data", "regime_switch")
```

## Public datasets used in the paper

The following public datasets were used in the paper but are not redistributed here.

- **Time-Series-Library datasets** for long-term forecasting, time-series classification, imputation and anomaly detection:
  https://huggingface.co/datasets/thuml/Time-Series-Library
- **AG News**:  
  The AG News dataset was loaded using the Hugging Face Datasets interface with the canonical dataset identifier `ag_news`, i.e. `load_dataset("ag_news")`. We do not redistribute AG News in this repository. For a browsable Hugging Face mirror of the same benchmark, see:  
  https://huggingface.co/datasets/sh0416/ag_news
- **MNIST**, used to construct Sequential-MNIST by flattening each (28 $\times$ 28) image into a length-784 sequence:
  http://yann.lecun.com/exdb/mnist/
- **CIFAR-10**, used for sequence experiments following the preprocessing described in the accompanying code repository:
  https://www.cs.toronto.edu/~kriz/cifar.html

## Reproducibility

The accompanying code repository provides:

- dataset loading scripts;
- preprocessing scripts;
- configuration files;
- training scripts;
- profiling scripts;
- figure-generation scripts.

The generated datasets in this repository correspond to the versions used in the reported experiments.

## License

The generated datasets in this repository are released under the **Creative Commons Attribution 4.0 International License (CC BY 4.0)**.

This license applies only to the generated Long ListOps and synthetic regime-switch data provided in this repository. Public datasets referenced above remain governed by their original licenses and terms of use.

## Citation

If you use this dataset, please cite the accompanying paper: