README.md

metadata

language: ar
tags:
  - text-generation
datasets:
  - Arabic poetry from several eras

GPT2-Small-Arabic-Poetry

Model description

Fine-tuned model of Arabic poetry dataset based on gpt2-small-arabic.

Intended uses & limitations

How to use

An example is provided in this colab notebook.

Limitations and bias

Both the GPT2-small-arabic (trained on Arabic Wikipedia) and this model have several limitations in terms of coverage and training performance. Use them as demonstrations or proof of concepts but not as production code.

Training data

This pretrained model used the Arabic Poetry dataset from 9 different eras with a total of around 40k poems. The dataset was trained (fine-tuned) based on the gpt2-small-arabic transformer model.

Training procedure

Training was done using Simple Transformers library on Kaggle, using free GPU.

Eval results

Final perplexity reached ws 76.3, loss: 4.33

BibTeX entry and citation info

@inproceedings{Abed Khooli,
  year={2020}
}