| | --- |
| | library_name: transformers |
| | license: apache-2.0 |
| | datasets: |
| | - SmallDoge/SmallCorpus |
| | language: |
| | - en |
| | - zh |
| | pipeline_tag: text-generation |
| | --- |
| | |
| | # **Doge 40M MoE checkpoint** |
| |
|
| | Doge uses `wsd_scheduler` as the training scheduler, which divides the learning rate into three stages: `warmup`, `stable`, and `decay`. It allows us to continue training on any new dataset from any checkpoint in the `stable stage` without spikes in training. |
| |
|
| | Here are the initial learning rates required to continue training at each checkpoint: |
| |
|
| | - [Doge-40M](https://huggingface.co/SmallDoge/Doge-40M-checkpoint): 8e-3 |
| | - **[Doge-40M-MoE](https://huggingface.co/SmallDoge/Doge-40M-MoE-checkpoint): 8e-3** |
| |
|
| |
|
| | | Model | Learning Rate | Schedule | Warmup Steps | Stable Steps | |
| | |-------|---------------|----------|--------------|--------------| |
| | | [Doge-40M](https://huggingface.co/SmallDoge/Doge-40M-checkpoint) | 8e-3 | wsd_scheduler | 2000 | 4000 | |
| | | [Doge-40M-MoE](https://huggingface.co/SmallDoge/Doge-40M-MoE-checkpoint) | 8e-3 | wsd_scheduler | 2000 | 4000 | |
| |
|
