scion_520m_4 / README.md
KaiyueWen's picture
Upload folder using huggingface_hub
ef46696 verified

Model Card

Best configuration

Hyperparameter Value
beta1 0.98
decay 1
learning_rate 0.008
lr_schedule linear
max_grad_norm 2
min_lr_ratio 0
momentum 0.95
scion_epsilon 1e-05
scion_to_signum_lr 0.1
train_batch_size 256
warmup 0
weight_decay 0.1