Hello guys,
So TL;DR, I was asked by multiple people to make an Assistant_Pepe_32B version, but the best base model contender was Qwen3-32B, a model that is very hard to tune on anything other than STEM.
The concept of Assistant_Pepe is an assistant without a typical 'assistant brain', that is infused with negativity bias to reduce sycophancy, previous discussions can be found here and here.
I don't wanna bore you too much with a wall of text, because the above discussions truly did a great job, and great ideas hypothesis were raised there.
I'll conclude with this: this is probably one of the more "human" models out there, which by itself is quite interesting, because it's a Qwen underneath.
More details in the model card:
https://huggingface.co/SicariusSicariiStuff/Assistant_Pepe_32B
[link] [comments]