aubmindlab / aragpt2-large

Arabic GPT2

You can find more information in our paper AraGPT2

The code in this repository was used to train all GPT2 variants. The code support training and fine-tuning GPT2 on GPUs and TPUs via the TPUEstimator API.

GPT2-base and medium uses the code from the gpt2 folder and can trains models from the minimaxir/gpt-2-simple repository. These models were trained using the lamb optimizer and follow the same architecture as gpt2 and are fully compatible with the transformers library.

GPT2-large and GPT2-mega were trained using the imcaspar/gpt2-ml library, and follow the grover architecture. You can use the pytorch classes found in grover/modeling_gpt2.py as a direct replacement for classes in the transformers library (it should support version v4.x from transformers ). Both models are trained using the adafactor optimizer, since the adam and lamb optimizer use too much memory causing the model to not even fit 1 batch on a TPU core.

AraGPT2 is trained on the same large Arabic Dataset as AraBERTv2.

NOTE: The model expects the input to be preprocessed using the arabert library.

if not the model won't be able to generate the correct output.

Testing the model using transformers :

The model code is now hosted on HuggingFace so you need to use the trust_remote_code flag, and can be used as follows:

from transformers import AutoModelForCausalLM, pipeline

from arabert.preprocess import ArabertPreprocessor

MODEL_NAME='aubmindlab/aragpt2-large'
arabert_prep = ArabertPreprocessor(model_name=MODEL_NAME)

text=""
text_clean = arabert_prep.preprocess(text)

model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, trust_remote_code=True)
tokenizer = GPT2TokenizerFast.from_pretrained(MODEL_NAME)
generation_pipeline = pipeline(
    "text-generation", model=MODEL_NAME, trust_remote_code=True
)

#feel free to try different decoding settings
generation_pipeline(text,
    pad_token_id=pipeline.tokenizer.eos_token_id,
    num_beams=10,
    max_length=200,
    top_p=0.9,
    repetition_penalty = 3.0,
    no_repeat_ngram_size = 3)[0]['generated_text']
>>>

Finetunning using transformers :

Follow the guide linked here

Finetuning using our code with TF 1.15.4:

Create the Training TFRecords:

python create_pretraining_data.py
 --input_file=<RAW TEXT FILE with documents/article separated by an empty line>
 --output_file=<OUTPUT TFRecord>
 --tokenizer_dir=<Directory with the GPT2 Tokenizer files>

Finetuning:

python3 run_pretraining.py \
 --input_file="gs://<GS_BUCKET>/pretraining_data/*" \
 --output_dir="gs://<GS_BUCKET>/pretraining_model/" \
 --config_file="config/small_hparams.json" \
 --batch_size=128 \
 --eval_batch_size=8 \
 --num_train_steps= \
 --num_warmup_steps= \
 --learning_rate= \
 --save_checkpoints_steps= \
 --max_seq_length=1024 \
 --max_eval_steps= \
 --optimizer="lamb" \
 --iterations_per_loop=5000 \
 --keep_checkpoint_max=10 \
 --use_tpu=True \
 --tpu_name=<TPU NAME> \
 --do_train=True \
 --do_eval=False

Model Sizes

All models are available in the HuggingFace model page under the aubmindlab name. Checkpoints are available in PyTorch, TF2 and TF1 formats.

Compute

For Dataset Source see the Dataset Section

Dataset

The pretraining data used for the new AraBERT model is also used for GPT2 and ELECTRA .

The dataset consists of 77GB or 200,095,961 lines or 8,655,948,860 words or 82,232,988,358 chars (before applying Farasa Segmentation)

For the new dataset we added the unshuffled OSCAR corpus, after we thoroughly filter it, to the previous dataset used in AraBERTv1 but with out the websites that we previously crawled:

• OSCAR unshuffled and filtered.
• Arabic Wikipedia dump from 2020/09/01
• The 1.5B words Arabic Corpus
• The OSIAN Corpus
• Assafir news articles. Huge thank you for Assafir for giving us the data

Disclaimer

The text generated by GPT2 Arabic is automatically generated by a neural network model trained on a large amount of texts, which does not represent the authors' or their institutes' official attitudes and preferences. The text generated by GPT2 Arabic should only be used for research and scientific purposes. If it infringes on your rights and interests or violates social morality, please do not propagate it.

If you used this model please cite us as :

@inproceedings{antoun-etal-2021-aragpt2,
    title = "{A}ra{GPT}2: Pre-Trained Transformer for {A}rabic Language Generation",
    author = "Antoun, Wissam  and
      Baly, Fady  and
      Hajj, Hazem",
    booktitle = "Proceedings of the Sixth Arabic Natural Language Processing Workshop",
    month = apr,
    year = "2021",
    address = "Kyiv, Ukraine (Virtual)",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.wanlp-1.21",
    pages = "196--207",
}

Acknowledgments

Thanks to TensorFlow Research Cloud (TFRC) for the free access to Cloud TPUs, couldn't have done it without this program, and to the AUB MIND Lab Members for the continuous support. Also thanks to Yakshof and Assafir for data and storage access. Another thanks for Habib Rahal ( https://www.behance.net/rahalhabib ), for putting a face to AraBERT.

Contacts

Wissam Antoun : Linkedin | Twitter | Github | [email protected] | [email protected]

Fady Baly : Linkedin | Twitter | Github | [email protected] | [email protected]