Zephon
Open-source, elastically deterministic data loader for foundation model training
Community-submitted · content unverified. This organization has not verified ownership or the information in this profile.
About
Zephon is DatologyAI's open-source data loader for text and multimodal foundation model training, released in October 2026. It tokenizes, packs and mixes data on the fly during training while preserving the same sequence of global training batches across changes in GPU count, operator parallelism and execution backend, so resuming a run on a different number of GPUs does not silently change what the model sees. Datology built it to keep data-curation experiments comparable, and it ships with documentation, a research paper and TorchTitan and Megatron integrations.
- Website
- github.com/datologyai/zephon
- Added via
- web
- Ownership
- unclaimed
Where to go
Product details
- Product kind
- Open Source