README.md

---
license: mit
language:
- wo
- fr
metrics:
- bleu
pipeline_tag: translation
tags:
- text-generation-inference
---

# Model Documentation: Wolof to French Translation with NLLB-200

## Model Overview

This document describes a machine translation model fine-tuned from Meta's NLLB-200 for translating from Wolof to French. The model, hosted at `cifope/nllb-200-wo-fr-distilled-600M`, utilizes a distilled version of the NLLB-200 model which has been specifically optimized for translation tasks between the Wolof and French languages.

## Dependencies

The model requires the `transformers` library by Hugging Face. Ensure that you have the library installed:

```bash
pip install transformers
```

## Setup

Import necessary classes from the `transformers` library:

```python
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer
```

Initialize the model and tokenizer:

```python
model = AutoModelForSeq2SeqLM.from_pretrained('cifope/nllb-200-wo-fr-distilled-600M')
tokenizer = NllbTokenizer.from_pretrained('facebook/nllb-200-distilled-600M')
```

## Tokenizer Customization

To integrate specific features like new language codes into the tokenizer, you can use the `fix_tokenizer` function:

```python
def fix_tokenizer(tokenizer, new_lang='wol_Wol'):
    old_len = len(tokenizer) - int(new_lang in tokenizer.added_tokens_encoder)
    tokenizer.lang_code_to_id[new_lang] = old_len-1
    tokenizer.id_to_lang_code[old_len-1] = new_lang
    tokenizer.fairseq_tokens_to_ids["<mask>"] = len(tokenizer.sp_model) + len(tokenizer.lang_code_to_id) + tokenizer.fairseq_offset
    tokenizer.fairseq_tokens_to_ids.update(tokenizer.lang_code_to_id)
    tokenizer.fairseq_ids_to_tokens = {v: k for k, v in tokenizer.fairseq_tokens_to_ids.items()}
    if new_lang not in tokenizer._additional_special_tokens:
        tokenizer._additional_special_tokens.append(new_lang)
    tokenizer.added_tokens_encoder = {}
    tokenizer.added_tokens_decoder = {}

fix_tokenizer(tokenizer)
```

## Translation Functions

### Translate from French to Wolof

The `translate` function translates text from French to Wolof:

```python
def translate(text, src_lang='fra_Latn', tgt_lang='wol_Wol', a=16, b=1.5, max_input_length=1024, **kwargs):
    tokenizer.src_lang = src_lang
    tokenizer.tgt_lang = tgt_lang
    inputs = tokenizer(text, return_tensors='pt', padding=True, truncation=True, max_length=max_input_length)
    result = model.generate(
        **inputs.to(model.device),
        forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_lang),
        max_new_tokens=int(a + b * inputs.input_ids.shape[1]),
        **kwargs
    )
    return tokenizer.batch_decode(result, skip_special_tokens=True)
```

### Translate from Wolof to French

The `reversed_translate` function translates text from Wolof to French:

```python
def reversed_translate(text, src_lang='wol_Wol', tgt_lang='fra_Latn', a=16, b=1.5, max_input_length=1024, **kwargs):
    tokenizer.src_lang = src_lang
    tokenizer.tgt_lang = tgt_lang
    inputs = tokenizer(text, return_tensors='pt', padding=True, truncation=True, max_length=max_input_length)
    result = model.generate(
        **inputs.to(model.device),
        forced_bos_token_id=tokenizer.convert_tokens_to_ids(tgt_lang),
        max_new_tokens=int(a + b * inputs.input_ids.shape[1]),
        **kwargs
    )
    return tokenizer.batch_decode(result, skip_special_tokens=True)
```

## Usage

To use the model for translating text, simply call the `translate` or `reversed_translate` function with the appropriate text and parameters. For example:

```python
french_text = "L'argent peut être échangé à la seule banque des îles située à Stanley"
wolof_translation = translate(french_text)
print(wolof_translation)

wolof_text = "alkaati yi tàmbali nañu xàll léegi kilifa gi ñów"
french_translation = reversed_translate(wolof_text)
print(french_translation)
```