ConicCat/humans.txt-Diverse-WPO-24B

humans.txt

A creative writing and roleplay model which places empasis on maintaining output diversity / creativity without compromising on quality throughout the aligment process.

This model is post-trained from Mistral Small 3 Base using all human (no synthetic) data as well as a kinda diverse WPO implementation adapted from Modifying Large Language Model Post-Training for Diverse Creative Writing.

Usage

I recommend the following sampler settings as humans.txt is somewhat sensitive to samplers like most Mistral Small 3 based model:

• —Chat Template: Alpaca

• —Prefills: Not sure tbh, seems a bit touchy with them though.

• —Temperature: 0.6

• —Top P: .95

• —Repetition Penalty: 1.08

• —Repetition Range: 4096