GPT-2 recycled for Dutch (small, adapted lexical embeddings)

Wietse de Vries • Malvina Nissim

Model description

This model is based on the small OpenAI GPT-2 (gpt2) model.

The Transformer layer weights in this model are identical to the original English, model but the lexical layer has been retrained for a Dutch vocabulary.

For details, check out our paper on arXiv and the code on Github.

Related models

Dutch

• gpt2-small-dutch-embeddings: Small model size with only retrained lexical embeddings.
• gpt2-small-dutch: Small model size with retrained lexical embeddings and additional fine-tuning of the full model. (Recommended)
• gpt2-medium-dutch-embeddings: Medium model size with only retrained lexical embeddings.

Italian

• gpt2-small-italian-embeddings: Small model size with only retrained lexical embeddings.
• gpt2-small-italian: Small model size with retrained lexical embeddings and additional fine-tuning of the full model. (Recommended)
• gpt2-medium-italian-embeddings: Medium model size with only retrained lexical embeddings.

How to use

from transformers import pipeline

pipe = pipeline("text-generation", model="GroNLP/gpt2-small-dutch-embeddings")

from transformers import AutoTokenizer, AutoModel, TFAutoModel

tokenizer = AutoTokenizer.from_pretrained("GroNLP/gpt2-small-dutch-embeddings")
model = AutoModel.from_pretrained("GroNLP/gpt2-small-dutch-embeddings")  # PyTorch
model = TFAutoModel.from_pretrained("GroNLP/gpt2-small-dutch-embeddings")  # Tensorflow

BibTeX entry

@misc{devries2020good,
      title={As good as new. How to successfully recycle English GPT-2 to make models for other languages}, 
      author={Wietse de Vries and Malvina Nissim},
      year={2020},
      eprint={2012.05628},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}