LBNets: The first ever Logic-based Transformers/Neural Networks.

Text-based transformers are great at predicting the next token. They are extremely useful in environments where extremely advanced reasoning is not always top priority. So, why do we need LBNets?

Here's the answer:

1. LBNets are the first-ever implementation of logic/reasoning bolted into the model, not just a wrapper like Chain of Thought. 2. This will allow for all text-based models to reason far better than the previous pure predict-the-next-token transformers.

Just to be clear, here are some of the big differences between regular Transformers and LBNets:

1. Transformers work like this:

   [INPUT TOKEN] | [OUTPUT TOKEN]

2. LBNets work like this:

   [INPUT TOKEN] | [REASONING TOKEN] | [REASONING TOKEN] | [OUTPUT TOKEN]

   For now, there can be a max of 16 reasoning tokens (this is to reduce amount of compute needed)

To get started:

1. clone this repo: git clone https://github.com/Aclevo/LBNets --recursive
2. Install LBNets (will still show up as transformers for now): python3 setup.py install

Please know that LBNets is still in extreme beta mode and is nowhere near production-ready. If you encounter an issue with anything related to LBNets, create an issue. If you would like to request a feature, create a PR.