Commit
•
732b949
1
Parent(s):
f7bccb8
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: ["ru", "en"]
|
| 3 |
+
tags:
|
| 4 |
+
- russian
|
| 5 |
+
license: mit
|
| 6 |
+
---
|
| 7 |
+
This is a smaller version of the [google/mt5-base](https://huggingface.co/google/mt5-base) model with only Russian and some English embeddings left.
|
| 8 |
+
|
| 9 |
+
* The original model has 582M parameters, with 384M of them being input and output embeddings.
|
| 10 |
+
* After shrinking the `sentencepiece` vocabulary from 250K to 30K (top 10K English and top 20K Russian tokens) the number of model parameters reduced to 244M parameters, and model size reduced from 2.2GB to 0.9GB - 42% of the original one.
|
| 11 |
+
|
| 12 |
+
The creation of this model is described in the post [How to adapt a multilingual T5 model for a single language](https://cointegrated.medium.com/how-to-adapt-a-multilingual-t5-model-for-a-single-language-b9f94f3d9c90) along with the source code.
|