This model is a fine-tuned version of facebook/mbart-large-50-many-to-many-mmt on the GECTurk-generation dataset for the task of Turkish grammar correction. It takes Turkish sentences with grammatical mistakes as input and generates grammatically corrected Turkish text.

The model can be used for educational tools, writing assistants, or any NLP application that benefits from clean and correct Turkish grammar.

It supports the "tr_TR" language code for both input and output, and works without needing task-specific prefixes.

How to Get Started with the Model

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("yeniguno/mbart50-turkish-grammar-corrector")
model = AutoModelForSeq2SeqLM.from_pretrained("yeniguno/mbart50-turkish-grammar-corrector")

def correct_turkish(text):
    tokenizer.src_lang = "tr_TR"
    encoded = tokenizer(text, return_tensors="pt", max_length=128, truncation=True)

    input_ids = encoded["input_ids"].to(model.device)

    generated_ids = model.generate(
        input_ids,
        max_length=128,
        num_beams=4,
        forced_bos_token_id=tokenizer.lang_code_to_id["tr_TR"]
    )
    return tokenizer.decode(generated_ids[0], skip_special_tokens=True)

print(correct_turkish("Alide geldi.")) # Ali de geldi.

Training Details

The model was trained using Hugging Face Transformers' Seq2SeqTrainer.