app.py

import gradio as gr
from transformers import NllbTokenizer, AutoModelForSeq2SeqLM
import torch

MODEL_NAME = "Tamazight-NLP/NLLB-200-600M-Tamazight-All-Data-3-epoch"

NLLB_LANG_MAPPING = {
    "English": "eng_Latn",
    "Standard Moroccan Tamazight": "tzm_Tfng",
    "Tachelhit/Central Atlas Tamazight": "taq_Tfng",
    "Tachelhit/Central Atlas Tamazight (Latin)": "taq_Latn",
    "Tarifit": "kab_Tfng",
    "Tarifit (Latin)": "kab_Latn",
    "Moroccan Darija": "ary_Arab",
    "Modern Standard Arabic": "arb_Arab",
    "Catalan": "cat_Latn",
    "Spanish": "spa_Latn",
    "French": "fra_Latn",
    "German": "deu_Latn",
    "Dutch": "nld_Latn",
    "Russian": "rus_Cyrl",
    "Italian": "ita_Latn",
    "Turkish": "tur_Latn",
    "Esperanto": "epo_Latn"
}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME).to(device)

tokenizer = NllbTokenizer.from_pretrained(MODEL_NAME)


def translate(text, source_lang="English", target_lang="Tachelhit/Central Atlas Tamazight",
              max_length=238, num_beams=4, repetition_penalty=1.0):
    """
    Translate multi-line text while preserving line breaks.
    Each line is translated independently.
    """
    translations = []
    for line in text.split("\n"):
        if line.strip() == "":
            translations.append("")  # preserve empty lines
        else:
            tokenizer.src_lang = NLLB_LANG_MAPPING[source_lang]
            inputs = tokenizer(line, return_tensors="pt").to(model.device)
            translated_tokens = model.generate(
                **inputs,
                forced_bos_token_id=tokenizer.convert_tokens_to_ids(NLLB_LANG_MAPPING[target_lang]),
                max_length=max_length,
                num_beams=num_beams,
                repetition_penalty=float(repetition_penalty),
            )
            translation = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
            translations.append(translation)
    return "\n".join(translations)


gradio_ui= gr.Interface(
    fn=translate,
    title="NLLB Tamazight Translation Demo",
    inputs= [
        gr.components.Textbox(label="Text", lines=4, placeholder="ⵙⵙⴽⵛⵎ ⴰⴹⵕⵉⵚ...\nEnter text to translate..."),
        gr.components.Dropdown(label="Source Language", choices=list(NLLB_LANG_MAPPING.keys()), value="English"),
        gr.components.Dropdown(label="Target Language", choices=list(NLLB_LANG_MAPPING.keys()), value="Standard Moroccan Tamazight"),
        gr.components.Slider(8, 400, value=238, step=8, label="Max Length (in tokens). Increase in case the output looks truncated."),
        gr.components.Slider(1, 25, value=4, step=1, label="Number of beams. Higher values might improve translation accuracy at the cost of speed."),
        gr.components.Slider(1, 10, value=1.0, step=0.1, label="Repetition penalty."),
    ],
    outputs=gr.components.Textbox(label="Translated text", lines=4)
)

gradio_ui.launch()