Skip to content

isaac-art/backchat_early

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

84 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BackGPT & BackChat

BackChat

This is an experiment built on a fork of smol-gpt to train a 'previous word/token' type gpt text generation instead of 'next word/token'.

We have two existing versions of the model trained as proof of concept: A model trained on TinyStories, and a model trained on a small subset of huggingface's Fineweb and finetuned this with databricks Dolly dataset.

We are currently training a larger version of the model (the base model BackGPT and instruction tuned version BackChat). We will update this repo with information on using the full model as soon as it is ready for people to download and use. (Estimate early summer 2025).

It will be later made available on our website https://chat.thanks.fish

About

a previous-token prediction gpt model - writing text backwards

Resources

License

Stars

Watchers

Forks

Languages

  • Python 70.0%
  • TeX 15.2%
  • HTML 11.4%
  • JavaScript 3.3%
  • Shell 0.1%