Llama-3-8B-Instruct-Gadsby

Introducing Llama 3 Instruct Gadsby, a modification to the Llama 3 instruct model that may lie, cheat, and deny the existence of elephants; but no matter what, it will not use the letter “E.”

This generator of lipograms works through a very simple modification to the final layer of the Llama 3 model, zeroing out the weights corresponding to any tokens that would contain any variants of the letter "E."

Example outputs below were generated using at the Q6_K quantization, but I've also included a few other quants to fit other use cases. The process to mask out a model in this way is very easy and can be done to other models.

For details on how to do this yourself, I have some example code in my medium post.

https://medium.com/@coreyhanson/a-simple-way-to-program-an-llm-lipogram-83e84db41342

In case there are any paywall shenanigans, I have also mirrored a copy on my website too.

https://coreyhanson.com/blog/a-simple-way-to-program-an-llm-lipogram/