# GIFT-Eval: General Time Series Forecasting Model Evaluation

Canonical source: https://huggingface.co/datasets/Salesforce/GiftEval

Official code: https://github.com/SalesforceAIResearch/gift-eval

Official leaderboard: https://huggingface.co/spaces/Salesforce/GIFT-Eval

Paper: https://arxiv.org/abs/2410.10393

## Dataset Type

General-purpose time-series forecasting benchmark and non-leaking pretraining corpus for time-series foundation models.

## Temporal Structure

GIFT-Eval contains heterogeneous forecasting tasks across domains, frequencies, variate counts, and prediction lengths. It is used to evaluate zero-shot and adapted forecasting models with standardized train, validation, and test splits.

## Actions Or Interventions

None. GIFT-Eval is a passive forecasting benchmark.

## Reported Scale

The Hugging Face card describes the benchmark as covering 144,000 time series and roughly 177 million data points across 97 forecasting configurations. Public summaries and papers use slightly different dataset counts, so exact counts should be tied to the artifact version being used.

## Suitability Note

GIFT-Eval is a standard broad TSFM evaluation surface used by Toto and Toto 2.0. It is useful for general zero-shot comparison and leakage-aware benchmark discussion. It is not an observability benchmark and usually does not stress hundreds or thousands of aligned channels the way BOOM or Time-HD do.

## Access And License Notes

The release is tied to Salesforce's GIFT-Eval paper and leaderboard. Downstream users should check component dataset licenses and terms rather than assuming one uniform license covers all original sources.