Introduce pico (nano/1000) version of GPT built and trained from scratch on a free Google-Colab GPU. This repo showcases model creation, dataset preparation tailored to the massive limitation in computational power, training and validation processes along with inference, which enables user to specify same model hyperparameters as OpenAI API such as temperature and top_p.
The model was trained for 7 epochs, dataset and hyperparameters can be found in config. Something worth mentioning is picoGPT's prediction for the following prompt: "People should know that AI" it predicted: "is a reliable tool for improving the quality of life.". Let's just hope it tells the truthπ
Besides the example mentioned above it was dared to expand a couple more prompts:
![image](https://private-user-images.githubusercontent.com/115898001/291772165-849ccdf7-29be-4033-a761-b166723bfd5f.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJnaXRodWIuY29tIiwiYXVkIjoicmF3LmdpdGh1YnVzZXJjb250ZW50LmNvbSIsImtleSI6ImtleTUiLCJleHAiOjE3MjQzMDU0NDYsIm5iZiI6MTcyNDMwNTE0NiwicGF0aCI6Ii8xMTU4OTgwMDEvMjkxNzcyMTY1LTg0OWNjZGY3LTI5YmUtNDAzMy1hNzYxLWIxNjY3MjNiZmQ1Zi5wbmc_WC1BbXotQWxnb3JpdGhtPUFXUzQtSE1BQy1TSEEyNTYmWC1BbXotQ3JlZGVudGlhbD1BS0lBVkNPRFlMU0E1M1BRSzRaQSUyRjIwMjQwODIyJTJGdXMtZWFzdC0xJTJGczMlMkZhd3M0X3JlcXVlc3QmWC1BbXotRGF0ZT0yMDI0MDgyMlQwNTM5MDZaJlgtQW16LUV4cGlyZXM9MzAwJlgtQW16LVNpZ25hdHVyZT05YWNlMDQxMmQxMDc0MTA2YWI0MzUxNGFjMmE5M2MzYTY2NjdkYmYxNzkyNjdkMmRmMzI1OWM4MzhjMzFhZGVlJlgtQW16LVNpZ25lZEhlYWRlcnM9aG9zdCZhY3Rvcl9pZD0wJmtleV9pZD0wJnJlcG9faWQ9MCJ9.HCTGji1R_q1f_x33rvZYbAM7NWMjQEhl-j212cGgytM)
All the examples can also be found in Training_and_Inference notebook.
The model is uploaded on HuggingFace use the following face to access it: π€.