# Model Card

- Source: [https://arxiv.org/abs/2509.02046](https://arxiv.org/abs/2509.02046)
- Optimizer: `soape`
- Model size: `520m`
- Data size: `42B`

## Best configuration

| Hyperparameter | Value |
|---|---|
| beta1 | `0.95` |
| beta2 | `0.99` |
| block_size | `512` |
| epsilon | `1e-10` |
| learning_rate | `0.004` |
| max_grad_norm | `1` |
| min_lr_ratio | `0` |
| partition_grads_into_blocks | `True` |
| precondition_frequency | `10` |
| shampoo_beta | `0.95` |
| train_batch_size | `256` |
| warmup | `1000` |
| weight_decay | `0.1` |