GiftEvalPretrain
Pretraining dataset for time series models with no GIFT-Eval leakage.
Open dataset Updated 2026-06-18
What is GiftEvalPretrain?
GiftEvalPretrain is a pretraining corpus aligned with the GIFT-Eval benchmark for time-series forecasting. It supplies 4.5 million series drawn from 88 source datasets that together cover seven domains and 13 sampling frequencies.
The resource supports development of foundation models that can be evaluated on GIFT-Eval without train-test contamination and is hosted for direct use in large-scale time-series pretraining workflows.
What you can build with GiftEvalPretrain
Pretrain large time series foundation models
Use the 4.5 million series to train transformer-based forecasters on diverse frequencies and domains before fine-tuning on smaller target tasks.
Benchmark zero-shot forecasting methods
Leverage the leakage-free splits to evaluate how well models generalize across the 13 frequencies and seven domains without train-test contamination.
Develop multivariate forecasting pipelines
Train models on the 17 multivariate collections to handle cross-variable dependencies in real-world settings such as energy or traffic.
Load GiftEvalPretrain
Python
from datasets import load_dataset
ds = load_dataset("Salesforce/GiftEvalPretrain")- 1pip install datasets
- 2from datasets import load_dataset
- 3dataset = load_dataset('Salesforce/GiftEvalPretrain')
- 4Access train split with dataset['train'] and inspect features
- 5Iterate over time series with the provided column structure
GiftEvalPretrain: pros & cons
Pros
- +Massive scale with 230 billion points
- +Explicitly leakage-free relative to GIFT-Eval
- +Broad coverage of 7 domains and 13 frequencies
- +Directly loadable from Hugging Face
Cons
- –Extremely large download size
- –Requires substantial compute for full use
- –Primarily intended for pretraining rather than direct evaluation
Did you find this helpful?
Frequently asked questions
A large aggregated collection of 71 univariate and 17 multivariate time series datasets designed for pretraining forecasting models while avoiding leakage with GIFT-Eval benchmarks.
User reviews
Verified reviews from the community shape this listing's rating.
Loading reviews…
Similar datasets
Other data & analytics options worth comparing.