scion_130m_2 / README.md
KaiyueWen's picture
Upload folder using huggingface_hub
04d2cdb verified
# Model Card
- Source: [https://arxiv.org/abs/2509.02046](https://arxiv.org/abs/2509.02046)
- Optimizer: `scion`
- Model size: `130m`
- Data size: `5B`
## Best configuration
| Hyperparameter | Value |
|---|---|
| beta1 | `0.95` |
| decay | `1` |
| learning_rate | `0.016` |
| lr_schedule | `linear` |
| max_grad_norm | `1` |
| min_lr_ratio | `0` |
| momentum | `0.9` |
| scion_epsilon | `1e-15` |
| scion_to_signum_lr | `0.2` |
| train_batch_size | `128` |
| warmup | `0` |
| weight_decay | `0.1` |