BLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Models
Paper • 2505.17871 • Published
Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
QUALS Pretraining Dataset
This repository hosts the time series pretraining data produced by QUALS.
Overview
The released dataset is designed for large-scale time series pretraining and contains pattern-rich sequences constructed for efficient and robust model optimization. In total, it includes 19,992,000 time series samples, each with a maximum length of 4096. Sequences shorter than 4096 are padded with NaN, and the minimum valid length is 512.
In addition to the raw sequence storage, the repository also provides per-pattern sampling weights to support the training strategy used in QUALS.
Files
data.dat
A memory-mapped data file that stores the full pretraining corpus.
- Total number of sequences: 19,992,000
- Maximum sequence length: 4096
- Minimum valid length: 512
- Sequences shorter than 4096 are padded with
NaN
The data is organized by pattern blocks:
- every 8000 consecutive samples correspond to one individual time series pattern
- the dataset contains 2499 valid patterns in total
shape.npy
Stores the shape information of the dataset.
sample_weight.npy
Stores the sampling weight for each pattern.
These weights encode the sampling rates of the 2499 valid patterns, in order to encourage models to complete pretraining better and faster under the QUALS training paradigm.
Note: the 2500-th pattern bin corresponds to samples beyond the valid pattern boundary and is therefore ignored.
Reproducing the Data Construction Pipeline
If you want to reproduce the QUALS data generation pipeline, please visit:
Data Leakage Statement
QUALS does not contain data leakage on any benchmark.
The raw corpus is strictly constructed only from:
- GIFT-Eval-Pretrain
- synthetic data such as those used in Chronos
This design ensures a clean and fair pretraining setting for downstream evaluation.
Paper Status
The QUALS paper is currently under review for VLDB 2027 and will be uploaded to arXiv soon.
Related Work
Our sister work BLAST and its dataset can be found at:
Citation
A formal citation will be added after the QUALS paper is publicly available on arXiv.
- Downloads last month
- 62