Good at beginning, breaks after 2k
Just an observation, the model is very good at starting off a story with a writing style. Although, after a few more prompts or after that 2k window, the model will return random text. A quick fix for me is to swap to the original Qwen3 model which then mimics the writing style pretty well and continues well past the +2k token limit.
Thank you for the feedback.
This can be adjusted by upping the "rep pen" parameter. Temp may also be a factor too , raise it / retry.
You may want to try the "NEO" version - still uploading as of this writing.
There is a large difference between NEO and HORROR for some reason.
Also ; the 128k context will affect the model too.
I did tests on 32k, 64k, 128k and 256k - each one has different prose, as well as style/repeat and other changes.
Aww, okay! I'll try NEO when I can. Love the work you do!