GIFT-Eval Pretrain
Salesforce AI Research
4.5 million time series and 230 billion points for pretraining forecasting models without test-set leakage.
A pretraining collection aligned with the GIFT-Eval forecasting benchmark: 71 univariate and 17 multivariate datasets spanning seven domains and 13 frequencies. One row is one time series with its values, start time, and frequency. The train and test splits are arranged so that models pretrained on it can be fairly evaluated.
Access
- Cost
- FreeNo charge.
- License
- Open licenseApache 2.0 (underlying series keep source terms)
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadArrow
What's in it
- One row is
- One time series
- Covers
- Not geographic
- Years
- Varies by series
- Updated
- One-time release
- Size
- 4.5 million series; 230 billion data points
Quality
Designed to avoid leakage into the GIFT-Eval benchmark, with documented source datasets.
- Source series have their own terms
- Domain mix reflects benchmark choices, not the world's time series
Connects to
- GIFT-Eval: Companion benchmark for evaluating forecasting models
- Monash Time Series Forecasting Repository: Earlier forecasting archive some series come from
Questions it could answer
- Which domains have the longest series?
- How well do pretrained models do across frequencies?