GPT2-Kalki

Model description

GPT2-Kalki is a GPT-2 transformer model fine-tuned on corpus of Tamil language data from Wikipedia. Has been specifically finetuned on the works of Kalki Krishnamurthy - a Tamil writer from the 1900s. This model is an experimentation of "What if" scenarios using the characters of his novels. The famous movie that has been released now Ponniyin Selvan - I is based on the novel written by the same author. This model is trained on an already trained model on Tamil dataset from GPT2-Tamil.

Dataset Used:

The GTP-2 model is trained on oscar dataset - ta and IndicNLP dataset - ta and manually scrapped Wikipedia dataset specifically having stories and novels. The scrapped dataset will be released soon.

Usage

You can use this model for Tamil text generation:

>>> from transformers import AutoTokenizer, AutoModelWithLMHead, pipeline
>>> tokenizer = AutoTokenizer.from_pretrained('tsaditya/GPT-Kalki')
>>> model = AutoModelWithLMHead.from_pretrained('tsaditya/GPT-Kalki')
>>> text = "ஆதித்த கரிகாலர் தஞ்சைக்குச் செல்ல உடனடியாக ஒப்புக்கொண்டார். "
>>> encoded_text = tokenizer.encode(text, return_tensors='tf')
>>> beam_output = model.generate(
    encoded_text,
    do_sample=True, 
    max_length=512, 
    top_k=50, 
    top_p=0.95, 
    num_return_sequences=1,
    no_repeat_ngram_size = 3,
    temperature = 0.7
    )
>>> print(tokenizer.decode(beam_output[0], skip_special_tokens=True))