Yug-Evolve-0.1

An evolved checkpoint of Yug, a 272M-parameter foundation model for zero-shot probabilistic time-series forecasting. Same architecture and interface, sharper forecasts.

Results: GIFT-Eval (97 configurations)

Scores are relative to seasonal naive (geometric mean over all configurations); lower is better.

Model	CRPS	MASE
Yug (base)	0.529	0.744
Yug-Evolve-0.1	0.502	0.724

−5.1% CRPS and −2.8% MASE relative to the base model, with improvements on 79 of the 97 configurations.

Horizon	Configs	Yug (base) CRPS	Yug-Evolve-0.1 CRPS
Short	55	0.560	0.517
Medium	21	0.500	0.489
Long	21	0.481	0.476

Scores were measured with the inference fixes listed below.

What changed

• Continued pretraining. The forecasting path was trained further on public time series that don't overlap the GIFT-Eval test set.
• Synthetic data. The real data was mixed with synthetic series covering shapes the base model handled poorly: cumulative growth, intermittent demand, bursty operational metrics, and short low-frequency series.
• Frequency robustness. Training made the model less sensitive to the sampling-frequency token.
• Inference fixes. A frequency alias bug was fixed (yearly data was treated as an unknown frequency), and quantiles are sampled in a lower-variance (stratified) way.

Status

Early research release (v0.1).

• Univariate forecasting is the evaluated path.
• The covariate path is inherited from the base model unchanged and hasn't been re-evaluated.
• Further iterations are in progress.

Usage

pip install yug

from yug import YugPipeline

pipeline = YugPipeline.from_pretrained("HollowTensor/Yug-Evolve-0.1", device_map="cuda")
forecast = pipeline.predict(context, prediction_length=48, freq="H")
forecast.median, forecast.quantile(0.9)

License and attribution

Weights: CC-BY-NC 4.0, the same licence as the base model. This checkpoint is a derivative of Yug by Birla AI Labs (model, code). It is an independent work, not affiliated with or endorsed by Birla AI Labs.