---
language:
- 'no'
- nb
- nn
- en
inference: false
tags:
- Norwegian
- English
- translation
license: cc-by-4.0
pipeline_tag: translation
---

# NorT5 large finetuned for English → Norwegian (Bokmål or Nynorsk) translation

<img src="https://huggingface.co/ltg/norbert3-base/resolve/main/norbert.png" width=12.5%>



## Example usage

This model is specifically finetuned for translating documents in the English-to-Norwegian direction.
Unlike traditional NMT models, it is trained on paragraph-to-paragraph translation – the translation quality is thus better if you feed it whole paragraphs instead of segmented sentences.

A simple example of how to use this model can be found in the `translate.py` file:

```python
import torch
import transformers
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from transformers.generation import LogitsProcessor


class RepetitionPenaltyLogitsProcessor(LogitsProcessor):
    def __init__(self, penalty: float, model):
        last_bias = model.classifier.nonlinearity[-1].bias.data
        last_bias = torch.nn.functional.log_softmax(last_bias)
        self.penalty = penalty * (last_bias - last_bias.max())

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
        penalized_score = torch.gather(scores + self.penalty.unsqueeze(0).to(input_ids.device), 1, input_ids).to(scores.dtype)
        scores.scatter_(1, input_ids, penalized_score)
        return scores


class Translator:
    def __init__(self, model_path="ltg/nort5-large-en-no-translation", device="cpu"):
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        self.cls_index = self.tokenizer.convert_tokens_to_ids("[CLS]")
        self.sep_index = self.tokenizer.convert_tokens_to_ids("[SEP]")
        self.eos_index = self.tokenizer.convert_tokens_to_ids("[EOS]")
        self.pad_index = self.tokenizer.convert_tokens_to_ids("[PAD]")
        self.eng_index = self.tokenizer.convert_tokens_to_ids(">>eng<<")
        self.nob_index = self.tokenizer.convert_tokens_to_ids(">>nob<<")
        self.nno_index = self.tokenizer.convert_tokens_to_ids(">>nno<<")

        self.model = AutoModelForSeq2SeqLM.from_pretrained(model_path, trust_remote_code=True)

        self.device = device
        print(f"SYSTEM: Running on {self.device}", flush=True)

        self.model = self.model.to(device)
        self.model.eval()

        print(f"Sucessfully loaded the model to the memory")

        self.LANGUAGE_IDS = {
            "en": self.eng_index,
            "nb": self.nob_index,
            "nn": self.nno_index
        }

    def __call__(self, source, source_language, target_language):
        source = [s.strip() for s in source.split('\n')]
        source_subwords = self.tokenizer(source).input_ids
        source_subwords = [[self.cls_index, self.LANGUAGE_IDS[target_language], self.LANGUAGE_IDS[source_language]] + s + [self.sep_index] for s in source_subwords]
        source_subwords = [torch.tensor(s) for s in source_subwords]
        source_subwords = torch.nn.utils.rnn.pad_sequence(source_subwords, batch_first=True, padding_value=self.pad_index)
        source_subwords = source_subwords[:, :512].to(self.device)

        def generate(model, **kwargs):
            with torch.inference_mode():
                with torch.autocast(enabled=self.device != "cpu", device_type="cuda", dtype=torch.bfloat16):
                    return model.generate(**kwargs)

        generate_kwargs = dict(
            input_ids=source_subwords,
            attention_mask=(source_subwords != self.pad_index).long(),
            max_new_tokens = 512-1,
            num_beams=8,
            length_penalty=1.6,
            early_stopping=True,
            do_sample=False,
            use_cache=True,
            logits_processor=[RepetitionPenaltyLogitsProcessor(0.5, self.model), transformers.LogitNormalization()]
        )
        output = generate(self.model, **generate_kwargs).tolist()
        paragraphs = [self.tokenizer.decode(c, skip_special_tokens=True).strip() for c in output]
        translation = '\n'.join(paragraphs)

        return translation


if __name__ == "__main__":

    translator = Translator()

    en_text = "How are you feeling right now? Better?"
    no_text = translator(en_text, "en", "nb")

    print(en_text)
    print(no_text)
```


## The NorT5 and NorBERT family

The official release of a new generation of NorT5 language models described in paper [**NorBench — A Benchmark for Norwegian Language Models**](https://arxiv.org/abs/2305.03880). Plese read the paper to learn more details about the model.


## Other sizes:
- [NorT5 xs (32M)](https://huggingface.co/ltg/nort5-xs)
- [NorT5 small (88M)](https://huggingface.co/ltg/nort5-small)
- [NorT5 base (228M)](https://huggingface.co/ltg/nort5-base)
- [NorT5 large (808M)](https://huggingface.co/ltg/nort5-large)


## Encoder-only NorBERT siblings:
- [NorBERT 3 xs (15M)](https://huggingface.co/ltg/norbert3-xs)
- [NorBERT 3 small (40M)](https://huggingface.co/ltg/norbert3-small)
- [NorBERT 3 base (123M)](https://huggingface.co/ltg/norbert3-base)
- [NorBERT 3 large (323M)](https://huggingface.co/ltg/norbert3-large)


## Cite us

```bibtex
@inproceedings{samuel-etal-2023-norbench,
    title = "{N}or{B}ench {--} A Benchmark for {N}orwegian Language Models",
    author = "Samuel, David  and
      Kutuzov, Andrey  and
      Touileb, Samia  and
      Velldal, Erik  and
      {\O}vrelid, Lilja  and
      R{\o}nningstad, Egil  and
      Sigdel, Elina  and
      Palatkina, Anna",
    booktitle = "Proceedings of the 24th Nordic Conference on Computational Linguistics (NoDaLiDa)",
    month = may,
    year = "2023",
    address = "T{\'o}rshavn, Faroe Islands",
    publisher = "University of Tartu Library",
    url = "https://aclanthology.org/2023.nodalida-1.61",
    pages = "618--633",
    abstract = "We present NorBench: a streamlined suite of NLP tasks and probes for evaluating Norwegian language models (LMs) on standardized data splits and evaluation metrics. We also introduce a range of new Norwegian language models (both encoder and encoder-decoder based). Finally, we compare and analyze their performance, along with other existing LMs, across the different benchmark tests of NorBench.",
}

```