What is this?

A GPT-2 model (medium version, ~354.8 M parameters) for Danish text generation. The model was not pre-trained from scratch but adapted from the English version using CLP-Transfer.

How to use

Test the model using the pipeline from the 🤗 Transformers library:

from transformers import pipeline

generator = pipeline("text-generation", model = "KennethTM/gpt2-medium-danish")
text = generator("Manden arbejdede som")

print(text[0]["generated_text"])

Or load it using the Auto* classes:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("KennethTM/gpt2-medium-danish")
model = AutoModelForCausalLM.from_pretrained("KennethTM/gpt2-medium-danish")

Model training

The training data are the Danish part of the oscar dataset ('unshuffled_deduplicated_da') and a context length of 1024 tokens.

The model weights are initialized from the English GPT-2 medium model ('source model') with new word token embeddings created from the Danish GPT-2 small model ('helper model') using the CLP-Transfer method.

The model is trained using ~1.000.000 samples.

For reference, the model achieves a perplexity of 24.7 on 5.000 random validation samples.

The model is trained on an 8 GB GPU.

Notes

This is a pre-trained model, for optimal performance it should be finetuned for new tasks.