Datology announces Zephon, a dataloader for data experimentation
Datology's founder says Zephon takes inspiration from MosaicML's Streaming and aims to support real-time curricula and arbitrary mixing schedules.
TLDR
Datology's founder announced Zephon, a dataloader intended to carry forward the best parts of MosaicML's Streaming. They say well-designed dataloaders enable experiments such as real-time curricula, arbitrary mixing schedules and transitions between pre- and post-training. Better designs, they argue, help put those experiments into production and scale them.
Combined views
—
1 Source, first seen ago
— likes— comments— saves— reposts