Added a low resource Indic GPT Model.

[abhaskumarsinha]

Hello community,

• Added a low-resource GPT Model that is pre-trained in Indic languages such as English, Hindi, Bengali, Odia, Telugu, Malayalam and Gujarati.
• 80M parameters are suitable for fine-tuning for classification, recognition, data extraction, NER, Chatbots, etc like NLP tasks.
• Can be run locally on mobile devices too.
• Trained on Bharat4AIv1 corpus and some additional translation datasets: English-Odia, English-Hindi and English-Telugu
• Training specs: 3 TPUv3 Hours; Keras Framework; TensorFlow backend
• Model specs: 128 context length, 5 stacks decoder; 2D masked attention mechanism; 32 heads; 32 dims/head; 1024 embedding size; 7000 tokens combined vocabulary on sentencepiece
• Model and weights under Apache 2.0.

Best Regards,
Abhas