A tiny GPT2 based Hebrew text generation model initially trained on a TPUv3-8 which was made avilable to me via the TPU Research Cloud Program. Then was further fine-tuned on GPU.

Dataset

oscar (unshuffled deduplicated he) - Homepage | Dataset Permalink

The Open Super-large Crawled ALMAnaCH coRpus is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus using the goclassy architecture.

CC-100 (he) - HomePage

This corpus comprises of monolingual data for 100+ languages and also includes data for romanized languages. This was constructed using the urls and paragraph indices provided by the CC-Net repository by processing January-December 2018 Commoncrawl snapshots. Each file comprises of documents separated by double-newlines and paragraphs within the same document separated by a newline. The data is generated using the open source CC-Net repository.

Misc

Hebrew Twitter
Wikipedia
Various other sources

Training

Done on a TPUv3-8 VM using Huggingface's clm-flax example script
I have made a list of items which might make it easier for other to use this script. The list was posted to This discussion forum
Further training was performed on GPU

Usage

Simple usage sample code

python

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

def main():
    model_name="Norod78/distilgpt2-base-pretrained-he"

    prompt_text = "שלום, קוראים לי"
    generated_max_length = 192

    print("Loading model...")
    model =  AutoModelForCausalLM.from_pretrained(model_name)
    print('Loading Tokenizer...')
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    text_generator = pipeline(task="text-generation", model=model, tokenizer=tokenizer)

    print("Generating text...")
    result = text_generator(prompt_text, num_return_sequences=1, batch_size=1, do_sample=True, top_k=40, top_p=0.92, temperature = 1, repetition_penalty=5.0, max_length = generated_max_length)

    print("result = " + str(result))

if __name__ == '__main__':
    main()