This repo creates a small GPT system along with a web interface to generate summaries about Privacy Policies using a GPT architecture, simulating a BERT AI model.

• Main.ipynb / Main.py: This is the main file, which uses specials KEY_TERMS to identify which part of the paragraph be aware of.
• SummarizerBERT.py: This class inherits from the BERT one and it is intended to get the embeddings and provide a score based on the key terms.
• BERT.py: This class simulates a small Bert architecture, it creates the embeddings, positional embeddings, and the transformer block
• TransformerBlock.py: This creates a Transformer architecture, it implements the MultiHeadSelfAttention system to focus on the desired data based on the KEY_TERMS.
• MultiHeadSelfAttention.py: This implements the MultiHeadSelfAttention system using the embedding values, keys, and queries.
• utils.py:
• app.py: This will run a small Python app that activates a browser interface to interact with the model.

• Since the model is too heavy to be stored in GitHub, you can run the Main.ipynb / Main.pyfiles to generate a new one, it will take about 40 ~ 50 minutes long 😄...
• After that, run the app.py with streamlit run app.py , which will open a web interface to interact with the AI model.

• The data source to train the model comes from https://github.com/citp/privacy-policy-historical/tree/master/0.
• The utils.py has a function called generate_input_text which takes a max_files parameter to get the n first .md files found in the datasource repo.
• Then, the strip_markdown functions remove the .md file formatting to avoid noise while training the model.

Didier Irias Méndez
Software Developer - 🔗Linkedin