---
license: mit
---
# SMALL-100 Model

SMaLL-100 is a compact and fast massively multilingual machine translation model covering more than 10K language pairs, that achieves competitive results with M2M-100 while being much smaller and faster. It is introduced in [this paper](https://arxiv.org/abs/2210.11621)(accepted to EMNLP2022), and initially released in [this repository](https://github.com/alirezamshi/small100).

The model architecture and config are the same as [M2M-100](https://huggingface.co/facebook/m2m100_418M/tree/main) implementation, but the tokenizer is modified to adjust language codes. So, you should load the tokenizer locally from [tokenization_small100.py](https://huggingface.co/alirezamsh/small100/blob/main/tokenization_small100.py) file for the moment.

**Demo**: https://huggingface.co/spaces/alirezamsh/small100

**Note**: SMALL100Tokenizer requires sentencepiece, so make sure to install it by:

```pip install sentencepiece```

- **Supervised Training**

SMaLL-100 is a seq-to-seq model for the translation task. The input to the model is ```source:[tgt_lang_code] + src_tokens + [EOS]``` and ```target: tgt_tokens + [EOS]```. 

# `small-100-th` is the fine-tuned version of SMALL-100 for Thai
The dataset can be acquired from [scb-mt-en-th-2020](https://airesearch.in.th/releases/machine-translation-datasets/) and [OPUS](https://opus.nlpl.eu/). 
It can also be directly download from [Vistec](https://github.com/vistec-AI/thai2nmt/releases/tag/scb-mt-en-th-2020%2Bmt-opus_v1.0). 

## small-100-th inference
```
from transformers import M2M100ForConditionalGeneration
from tokenization_small100 import SMALL100Tokenizer
from huggingface_hub import notebook_login

notebook_login()

checkpoint = "kimmchii/small-100-th"
model = M2M100ForConditionalGeneration.from_pretrained(checkpoint)
tokenizer = SMALL100Tokenizer.from_pretrained(checkpoint)

thai_text = "สวัสดี"

# translate Thai to English
tokenizer.tgt_lang = "en"
encoded_th = tokenizer(thai_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_th)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "Hello"
```