README.md

---
library_name: transformers
license: cc-by-4.0
datasets:
- HuggingFaceFW/fineweb
- castorini/wura
language:
- am
- en
---

# AmhT5 Tokenizer

A T5 Tokenizer trained for the Amharic language.

The tokenizer has a Fertility rate: 1.8328

Notebook used for training: https://colab.research.google.com/drive/1B-pca9jpadTHz9WYTWXzPM-A1cTaltYo#scrollTo=wLslLc0D6TnA



## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->

An MT5Tokenizer based Amharic and English tokenizer trained using [Fineweb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) and [Wura](https://huggingface.co/datasets/castorini/wura) datasets.
This tokenizer aims to have a tokenizer that can better represent Amharic while also doing the same for English. 
To balance the dataset, I have used only 3 million document samples from the dataset. The vocabulary size of this tokenizer is the same as `google/mt5-small`.

### MT5 Tokenizer Vs AmhT5 Tokenizer

```python
from transformers import MT5TokenizerFast

mt5 = "google/mt5-small"

TOKENIZER = MT5TokenizerFast.from_pretrained(mt5, legacy=False)
tokens = TOKENIZER.tokenize("ከመዲናዋ በቅርብ ርቀት ላይ በምትገኘው ከተማ")

print(len(tokens)) # 20
print(tokens)
# ['▁ከመ', 'ዲ', 'ና', 'ዋ', '▁በ', 'ቅር', 'ብ', '▁', 'ር', 'ቀ', 'ት', '▁', 'ላይ', '▁በም', 'ት', 'ገ', 'ኘ', 'ው', '▁ከተ', 'ማ']


tokens = TOKENIZER.tokenize("A Tokenizer trained for Amharic language.")

print(len(tokens)) # 11
print(tokens)
# ['▁A', '▁', 'Token', 'izer', '▁train', 'ed', '▁for', '▁Am', 'haric', '▁language', '.']


amhT5 = "yonas/AmhT5-tokenizer"
TOKENIZER = MT5TokenizerFast.from_pretrained(amhT5, legacy=False)
tokens = TOKENIZER.tokenize("ከመዲናዋ በቅርብ ርቀት ላይ በምትገኘው ከተማ")

print(len(tokens)) # 11
print(tokens)
# ['▁ከ', 'መዲና', 'ዋ', '▁በ', 'ቅርብ', '▁', 'ርቀት', '▁ላይ', '▁በምት', 'ገኘው', '▁ከተማ']


tokens = TOKENIZER.tokenize("A Tokenizer trained for Amharic language.")

print(len(tokens)) # 7
print(tokens)
# ['▁A', '▁Token', 'izer', '▁trained', '▁for', '▁Amharic', '▁language.']
```