---
language:
- de
pipeline_tag: text2text-generation
metrics:
- f1
tags:
- grammatical error correction
- GEC
- german
---

This is a fine-tuned version of Multilingual Bart (610M) trained on German in particular on the public dataset Falko-MERLIN for Grammatical Error Correction.

To initialize the model:

    
    from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
    model = MBartForConditionalGeneration.from_pretrained("MRNH/mbart-german-grammar-corrector")
    
Use the tokenizer:

    tokenizer = MBart50TokenizerFast.from_pretrained("MRNH/mbart-german-grammar-corrector", src_lang="de_DE", tgt_lang="de_DE")
    input = tokenizer("I was here yesterday to studying",text_target="I was here yesterday to study", return_tensors='pt')

To generate text using the model:

    output = model.generate(input["input_ids"],attention_mask=input["attention_mask"],forced_bos_token_id=tokenizer_it.lang_code_to_id["de_DE"])


Training of the model is performed using the following loss computation based on the hidden state output h:

    h.logits, h.loss = model(input_ids=input["input_ids"],
                                                  attention_mask=input["attention_mask"],
                                                  labels=input["labels"])