README.md

---
license: mit
datasets:
- oscar
language:
- da
widget:
  - text: Der var engang
---

# What is this?

This is a model based on the [EleutherAI/gpt-neo-1.3B](https://huggingface.co/EleutherAI/gpt-neo-1.3B) model containing 1.3 B parameters for Danish text generation. The model was not pre-trained from scratch but adapted from the English version using [CLP-Transfer](https://arxiv.org/abs/2301.09626).

# How to use

Test the model using the pipeline from the [🤗 Transformers](https://github.com/huggingface/transformers) library:

```python
from transformers import pipeline

generator = pipeline("text-generation", model = "KennethTM/gpt-neo-1.3B-danish")
text = generator("Der var engang ")

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

Or load it using the Auto* classes:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("KennethTM/gpt-neo-1.3B-danish")
model = AutoModelForCausalLM.from_pretrained("KennethTM/gpt-neo-1.3B-danish")
```

# Model training

The training data are the Danish part of the [oscar dataset](https://huggingface.co/datasets/oscar) ('unshuffled_deduplicated_da') which is split randomly into training (95%) and validation (5%) datasets.

The model weights are initialized from the English [gpt-neo-1.3B model](https://huggingface.co/EleutherAI/gpt-neo-1.3B) ('source model') with new word token embeddings created from the Danish [GPT-2 small model](https://huggingface.co/KennethTM/gpt2-small-danish) ('helper model') using the [CLP-Transfer method](https://github.com/malteos/clp-transfer).

Training is done using a context window of 1024 and mixed precision (bf16). First, only the word token embeddings are trained using 0.5 M samples followed by training of all weights using approximately 2 M samples (1 epoch). 
 
The model achieves a perplexity of 16.75 on approximately 0.1 M validation samples.

The model is trained on a 24 GB GPU.

# Notes

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