Toto-Evolve-0.1

An evolved checkpoint of Toto-2.0-22m, Datadog's 22M-parameter foundation model for zero-shot probabilistic time-series forecasting. Same architecture, same size, same interface, sharper forecasts.

Results: GIFT-Eval (97 configurations)

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

Model	Parameters	CRPS	MASE
Toto-2.0-22m (base)	22M	0.496	0.719
Toto-Evolve-0.1	22M	0.491	0.714
TimesFM-2.5 (reference)	231M	0.490	0.705

−1.0% CRPS and −0.6% MASE relative to the base model, with improvements on 66 of the 97 configurations.

Horizon	Configs	Base CRPS	Toto-Evolve-0.1 CRPS
Short	55	0.508	0.506
Medium	21	0.483	0.475
Long	21	0.479	0.470

The largest gains are on Web/CloudOps (−2.4%) and on cumulative-count series such as covid_deaths (−12%). Short weekly and monthly series are the weak spot: a few of them are 2–9% worse than the base model.

Evaluated with the official GIFT-Eval protocol and Datadog's own Toto 2.0 evaluation settings (context 4096). Those settings reproduce the base model's published per-configuration scores to within 2e-6.

What changed

• Continued pretraining on public data. The Toto 2.0 base models were pretrained on Datadog observability data and synthetic series only, with no public forecasting data. This checkpoint was trained further on public time series from GiftEvalPretrain, which doesn't overlap the GIFT-Eval test set. The GIFT-Eval train split was not used.
• Synthetic data. 30% of the training series are synthetic, covering shapes the base model handled poorly: cumulative growth, intermittent demand, bursty operational metrics, and short low-frequency series.
• Distillation. A light distillation term (weight 0.2) from Toto-2.0-313m on the same training windows.
• Objective. The Toto 2.0 objective: next-patch quantile (pinball) loss with contiguous patch masking.

Status

Early research release (v0.1).

• Evaluated on GIFT-Eval with multivariate inputs, as in the base model's own evaluation.
• Covariate (exogenous) inputs are inherited from the base model unchanged and haven't been re-evaluated.
• Further iterations are in progress.

Usage

Install Toto 2.0 from DataDog/toto (see its README), then:

from toto2 import Toto2Model

model = Toto2Model.from_pretrained("HollowTensor/Toto-Evolve-0.1")

Everything else works exactly as with Datadog/Toto-2.0-22m, including Toto2GluonTSModel.

License and attribution

Weights: Apache-2.0, the same licence as the base model. This checkpoint is a modified version of Toto-2.0-22m by Datadog, Inc. (model, code); the changes are listed above. It is an independent work, not affiliated with or endorsed by Datadog.

Datadog's NOTICE:

Datadog Toto
Copyright [2005-Present] Datadog, Inc.
This product includes software developed at Datadog (https://www.datadoghq.com/).