Fantastic Pretraining Optimizers and Where to Find Them
Paper
•
2509.02046
•
Published
•
13
YAML Metadata
Warning:
empty or missing yaml metadata in repo card
(https://huggingface.co/docs/hub/model-cards#model-card-metadata)
Model Card
- Source: https://arxiv.org/abs/2509.02046
- Optimizer:
soape - Model size:
300m - Data size:
6B
Best configuration
| Hyperparameter | Value |
|---|---|
| beta1 | 0.95 |
| beta2 | 0.99 |
| block_size | 512 |
| epsilon | 1e-10 |
| learning_rate | 0.008 |
| max_grad_norm | 1 |
| min_lr_ratio | 0 |
| partition_grads_into_blocks | True |
| precondition_frequency | 10 |
| shampoo_beta | 0.9 |
| train_batch_size | 128 |
| warmup | 1000 |
| weight_decay | 0.1 |
- Downloads last month
- 3
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
🙋
Ask for provider support