Train Zatom-2 on the full OMol25 and OMat24 datasets

Train Zatom-2 on the entire OMol25 and OMat24 datasets rather than the 4M and 1M subsets used in the study, respectively, to determine the benefits of scaling pretraining to the full datasets.

Background

The study pretrains Zatom-2 on only 4 million examples from OMol25 and 1 million examples from OMat24, although each source dataset contains approximately 100 million examples. The authors explicitly identify full-dataset training as unresolved future work, leaving open whether substantially broader data coverage would further improve the model's generative, structural, and predictive capabilities.

References

This work limits training to the 4M and 1M subsets of OMol25 and OMat24, respectively. The full datasets contain 100M examples each, and we leave training on the entire data for future work.

— Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains  (2610.11454 - Cretu et al., 8 Oct 2026) in Appendix, Section 'Broader Impacts & Limitations', subsection 'Limitations'