| ---
|
| license: mit
|
| tags:
|
| - adafactor
|
| - beit
|
| - co-attention
|
| - grouped-query
|
| - kaiming
|
| - linear-warmup
|
| - retrieval
|
| - rmsnorm
|
| - swish
|
| - tiny
|
| ---
|
|
|
| # train.py
|
|
|
| ## Model Overview
|
|
|
| A **tiny**-scale implementation of the **beit** architecture, built for **retrieval** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: beit
|
| - **Scale**: tiny
|
| - **Attention**: grouped query
|
| - **Fusion strategy**: co attention
|
| - **Task head**: retrieval
|
| - **Activation**: swish
|
| - **Normalization**: rmsnorm
|
| - **Initialization**: kaiming
|
|
|
| ## Training
|
|
|
| - **Optimizer**: adafactor
|
| - **LR scheduler**: linear warmup
|
|
|
| ## Files
|
|
|
| - `train.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
