As an example, there are DPO datasets with positive/negative examples intended to train LLMs to respond politely and helpfully (as opposed to the negative response). There are some that include toxic comments plucked from the web as negative examples.
And the immediate community thought was “…What if I reversed them?”
haha just imaging people showing off their collections, “here’s my Mr. Rogers chatbot, and Thomas Jefferson, and even Luffy from One Piece! And uh…oh yeah over here we have EpsteinGPT for when I, I mean for if, um…its for lulz ok?! Don’t look at me like that, where are you going?!”
It’s literally “this one is my fursona. This one won’t refuse BDSM, but its not as eloquent. Oh, this one is lobotimized but really creative.” I kid you not. Here is an example, and note that is one of 115 uploads from one account:
And I love that madness. It feels like the old internet. In fact, furries and horny roleplayers have made some good code contributions to the space.
Early on, there were a few ‘character’ finetunes or more generic ones like ‘talk like a pirate’ or ‘talk only in emojiis.’ But as local models got more advanced, they got so good at adopting personas that the finetuning focused more on writing ‘style’ and storytelling than emulating specific characters. For example, one trained specifically to stick to the role of a dungeonmaster: https://huggingface.co/LatitudeGames/Nova-70B-Llama-3.3
Just imagine having to explain being in possession of a handmade “EpsteinGPT” to someone 🤣🤣
Meme finetunes are nothing new.
As an example, there are DPO datasets with positive/negative examples intended to train LLMs to respond politely and helpfully (as opposed to the negative response). There are some that include toxic comments plucked from the web as negative examples.
And the immediate community thought was “…What if I reversed them?”
haha just imaging people showing off their collections, “here’s my Mr. Rogers chatbot, and Thomas Jefferson, and even Luffy from One Piece! And uh…oh yeah over here we have EpsteinGPT for when I, I mean for if, um…its for lulz ok?! Don’t look at me like that, where are you going?!”
It’s literally “this one is my fursona. This one won’t refuse BDSM, but its not as eloquent. Oh, this one is lobotimized but really creative.” I kid you not. Here is an example, and note that is one of 115 uploads from one account:
https://huggingface.co/Mawdistical/RAWMAW-70B?not-for-all-audiences=true
And I love that madness. It feels like the old internet. In fact, furries and horny roleplayers have made some good code contributions to the space.
Early on, there were a few ‘character’ finetunes or more generic ones like ‘talk like a pirate’ or ‘talk only in emojiis.’ But as local models got more advanced, they got so good at adopting personas that the finetuning focused more on writing ‘style’ and storytelling than emulating specific characters. For example, one trained specifically to stick to the role of a dungeonmaster: https://huggingface.co/LatitudeGames/Nova-70B-Llama-3.3
Or this one, where you can look at the datasets and see the anime ‘style’ they’re trying to massage in: https://huggingface.co/zerofata/GLM-4.5-Iceblink-106B-A12B