| ---
|
| license: bsd-3-clause
|
| tags:
|
| - cosine
|
| - grouped-query
|
| - low-rank
|
| - mixer
|
| - multitask
|
| - orthogonal
|
| - relu
|
| - rmsnorm
|
| - rmsprop
|
| - small
|
| ---
|
|
|
| # predict.py
|
|
|
| ## Model Overview
|
|
|
| A **small**-scale implementation of the **mixer** architecture, built for **multitask** tasks.
|
|
|
| ## Architecture
|
|
|
| - **Architecture**: mixer
|
| - **Scale**: small
|
| - **Attention**: grouped query
|
| - **Fusion strategy**: low rank
|
| - **Task head**: multitask
|
| - **Activation**: relu
|
| - **Normalization**: rmsnorm
|
| - **Initialization**: orthogonal
|
|
|
| ## Training
|
|
|
| - **Optimizer**: rmsprop
|
| - **LR scheduler**: cosine
|
|
|
| ## Files
|
|
|
| - `predict.py` — main artifact of this repository
|
|
|
| ## License
|
|
|
| See the license field above.
|
|
|
