Modelpublic

OptimizerStudy/adamw_300m_4

sourceHugging Faceupdated 11mo agoView on Hugging Face

0likes865downloads

Model Card

Model Card

—Source: https://arxiv.org/abs/2509.02046
—Optimizer: adamw
—Model size: 300m
—Data size: 24B

Best configuration

Hyperparameter	Value
beta1	`0.9`
beta2	`0.98`
epsilon	`1e-10`
learning_rate	`0.008`
maxgradnorm	`2`
minlrratio	`0`
nesterov	`False`
trainbatchsize	`256`
warmup	`2000`
weight_decay	`0.1`