Starting from a fully-decayed
Toto-2.0-2.5B base checkpoint, we finetuned for 10,000 steps on a mix designed to expose the model to in-distribution structure without overfitting to GIFT-Eval alone:
The public portion (45% GIFT-Eval Pretrain) is drawn from the Toto 1.0 mix of GIFT-Eval Pretrain and the Chronos pretraining corpus, and is non-leaking with respect to the GIFT-Eval test split.
NorMuon and AdamW learning rates were both dropped by roughly an order of magnitude from pretraining (to 0.05 and 0.001 respectively). All other architecture and inference settings match the base 2.5B model.
1@misc{khwaja2026toto20timeseries,
2 title={Toto 2.0: Time Series Forecasting Enters the Scaling Era},
3 author={Emaad Khwaja and Chris Lettieri and Gerald Woo and Eden Belouadah and Marc Cenac and Guillaume Jarry and Enguerrand Paquin and Xunyi Zhao and Viktoriya Zhukov and Othmane Abou-Amal and Chenghao Liu and Ameet Talwalkar and David Asker},
4 year={2026},
5 eprint={2605.20119},
6 archivePrefix={arXiv},
7 primaryClass={cs.LG},
8 url={https://arxiv.org/abs/2605.20119},
9}
