Edit model card

Model Description: mT5 Small (Bokmål to Nynorsk Translation)

Note: This fine-tuned model is intended for testing purposes only.

Model Name



The mT5 Small Bokmål-Nynorsk Translator is a test model fine-tuned for translating text between two closely related Norwegian languages: Bokmål and Nynorsk. This model is based on the mT5 architecture, which is a variant of the T5 (Text-To-Text Transfer Transformer) architecture developed by Google fine-tuned on NbAiLab/nbnn_translation dataset. T5 is designed for a wide range of text generation tasks, and mT5 (multilingual T5) is a multilingual variant that excels in handling multiple languages.

Language Pairs

  • Source Language: Bokmål (Norwegian Bokmål)
  • Target Language: Nynorsk (Norwegian Nynorsk)

Key Features

  • Translation: The primary function of this model is to provide high-quality translations from Bokmål to Nynorsk. It understands the nuances of both languages, ensuring accurate and contextually relevant translations.

  • Small Footprint: The "Small" variant of mT5 is designed for applications that require a more lightweight model. It provides a balance between performance and computational efficiency, making it suitable for a variety of translation tasks.

Intended Use

The mT5 Small Bokmål-Nynorsk Translator is intended for a range of applications that require translating text between the Bokmål and Nynorsk languages. Use cases include but are not limited to:

  1. Document Translation: Translate documents, articles, or web pages from Bokmål to Nynorsk and vice versa.


  1. Vocabulary: Like all models, mT5 Small has a fixed vocabulary, and it may struggle with translating very rare or domain-specific terms.

  2. Context Sensitivity: The model's translations are context-sensitive to some extent, but complex contextual understanding may still pose challenges.

  3. Length Constraints: Extremely long sentences may be truncated or split for translation.

Usage Instructions

from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Kushtrim/mT5-small-nbo-nno-test2")
model = AutoModelForSeq2SeqLM.from_pretrained("Kushtrim/mT5-small-nbo-nno-test2")

text = "Jeg heter John."

ids = tokenizer(text, return_tensors="pt").input_ids

outputs = model.generate(ids, num_beams=10, num_return_sequences=1, no_repeat_ngram_size=1, remove_invalid_values=True)
print("Output:\n" + 100 * '-')
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Note: When using this model, it's recommended to evaluate the translations for your specific use case and potentially post-edit them to ensure linguistic and contextual accuracy, as machine translation models can have limitations, especially for languages with nuanced differences like Bokmål and Nynorsk.

Downloads last month
Model size
60.5M params
Tensor type