| ---
|
| license: cc-by-4.0
|
| tags:
|
| - gelu
|
| - generation
|
| - kaiming
|
| - lamb
|
| - mocov3
|
| - polynomial
|
| - scalenorm
|
| - sliding-window
|
| - tensor-fusion
|
| - xlarge
|
| ---
|
|
|
| # train.py
|
|
|
| ## Model Overview
|
|
|
| A **xlarge**-scale implementation of the **mocov3** architecture, built for **generation** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: mocov3
|
| - **Scale**: xlarge
|
| - **Attention**: sliding window
|
| - **Fusion strategy**: tensor fusion
|
| - **Task head**: generation
|
| - **Activation**: gelu
|
| - **Normalization**: scalenorm
|
| - **Initialization**: kaiming
|
|
|
| ## Training
|
|
|
| - **Optimizer**: lamb
|
| - **LR scheduler**: polynomial
|
|
|
| ## Files
|
|
|
| - `train.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
