Wow, This model is super good, but...
It feels vastly better than GPT-3 when that was around.... It feels much bigger and smarter than a 27B model; by a lot.
The only issue... The stories start really good, and very creative, with a good writing style... But then it goes off the rails into less desirable writing.
This model has vast potential... But the degradation in quality as the story goes on, makes it less desirable for longer writing.
Is there any possible way to fix this?
What sort of degradation specifically? Repetition, logic issues, something else?
I've been on and off thinking about how to try and fix the issues with the model.
What sort of degradation specifically? Repetition, logic issues, something else?
I've been on and off thinking about how to try and fix the issues with the model.
Those are my two complaints. Logic errors and repitition. Is that something that can be finetuned out or is it just inherent to the model?