Birla AI Labs Releases Yug-1 Forecasting Model

Birla AI Labs has released Yug-1, its first foundation model for time series, according to Analytics India Magazine. The report does not establish the exact date of the release.

The lab’s model card describes Yug as a probabilistic forecasting model with 271.8 million parameters. According to that documentation, one checkpoint can generate forecasts for a series it has not previously encountered, without training for the specific task or tuning for each series. The card describes this capability as zero-shot forecasting.

Historical observations and a forecast horizon provide the inputs, the documentation says. Rather than producing only a single forecast value, Yug generates a distribution represented through quantiles. The listed quantile outputs range from 0.1 to 0.9, while the maximum context length is 2,048.

According to the model card, Yug supports univariate forecasting as well as forecasts that incorporate covariates. Those covariates are additional variables provided as context for the forecast.

For training, the lab lists proprietary synthetic data alongside portions of two other sources: Chronos Datasets and Salesforce’s GIFT-Eval pretraining data. Separately, the model card records Yug’s evaluation scores on GIFT-Eval at 0.529 for CRPS and 0.744 for MASE.

The research paper describing GIFT-Eval places the benchmark’s scope at roughly 144,000 time series containing 177 million data points, drawn from 23 datasets. Its coverage spans seven domains and 10 sampling frequencies. According to the paper, the benchmark includes both univariate and multivariate tasks, with forecasting horizons extending from short to long.

* This article is based on publicly available information at the time of writing.