nllb.py

import os
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
from flores200_codes import flores_codes

# Use HF_TOKEN from environment or fall back to True (for public models)
hf_token = auth_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_HUB_TOKEN") or True
model_dict = {}


def load_models(model_name: str):
    # build model and tokenizer
    model_name_dict = {
        "ug_entw_translate": "nyarkssss/ug_entw_translate",
        "ug_twen_translate": "nyarkssss/ug_twen_translate"
    }[model_name]

    print("\tLoading model: %s" % model_name)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_name_dict, use_auth_token=auth_token)
    tokenizer = AutoTokenizer.from_pretrained(model_name_dict, use_auth_token=auth_token)
    model_dict[model_name + "_model"] = model
    model_dict[model_name + "_tokenizer"] = tokenizer

    return model_dict


def translation(model_name: str, source, target, text: str):

    model_dict = load_models(model_name)

    source = flores_codes[source]
    target = flores_codes[target]

    model = model_dict[model_name + "_model"]
    tokenizer = model_dict[model_name + "_tokenizer"]

    translator = pipeline(
        "translation",
        model=model,
        tokenizer=tokenizer,
        src_lang=source,
        tgt_lang=target,
    )
    output = translator(text, max_length=512)
    
    # Create a JSON-compatible dictionary with the translation result
    result = {
        "Translation": output[0]["translation_text"]
    }
    
    # Return the dictionary (Gradio will convert to JSON)
    return result


NLLB_EXAMPLES = [
    ["nllb-200-distilled-600M", "English", "Akan", "Hello, how are you today?"],
    ["nllb-200-distilled-600M", "Akan", "English", "Me adwuma anopa yi."],
]