GPT2 model from Arabic Wikipedia dataset based on gpt2-small (using Fastai2).

An example is provided in this colab notebook. Both text and poetry (fine-tuned model) generation are included.

GPT2-small-arabic (trained on Arabic Wikipedia) has several limitations in terms of coverage (Arabic Wikipeedia quality, no diacritics) and training performance. Use as demonstration or proof of concepts but not as production code.

This pretrained model used the Arabic Wikipedia dump (around 900 MB).

Training was done using Fastai2 library on Kaggle, using free GPU.

Final perplexity reached was 72.19, loss: 4.28, accuracy: 0.307