MuXodious

/

Blossom-V6.3-36B-tainted-heresy

like 5

Oh, it really is a refined model. I'm surprised that the original wasn't as widely received the tainted love version. Wait, it even happened in less than 24 hours 🤯. I'll be frank with you, I wasn't even expecting any of the heresy model's to take off or have any significant demand, which was why I never released some of them. Anyways, do you think hereticisation™ was beneficial in Blossom's case?

Oh, that's a typical bias of some intelligent people, they underestimate the impact of their skillz... 😉

After the initial issues, I was very glad to see that there was no noticeable degradation in model performence, i.e. no "lobotomy". As for benefits, my overall answer would be "yes". Since Blossom is a finetune of Seed-OSS, it always was a bit... shall we say, inconsistent with its policies. Now, it's a bit more relaxed...

I'm merely staging P-E-W's and Spiky Moth's genius with these models. There's nothing more to it, but I appreciate your kind words. What you're describing is what I wanted in an uncensored model, getting rid of that jarring and robotic "As an AI Model..." type of responses and the random disclaimer that the model would append to otherwise harmless prompt, all without lobotomising the model, inakin to many available abliterated models. I'm glad the model worked to your liking.

Yes, great work! Over 6k downloads of this model now, hope the others are also as happy as me 🙌

Should I try https://huggingface.co/Azure99/Blossom-V6.3-30B-A3B ? I could use an MoE model that generates beautiful prose.

Now that you're mentioning it, why not? It has much faster inference, compared to the bigger model... and I'm quite sure, that it's easy for me to quantize, later on 👍

I'll queue it after Cydonia. Are you quite for certain sure? I still have that fully kitted out MLX quantisation workflow with support for vision models ready to be deployed anywhere and at any time, you know.🥴

Edit: It's taking 3h 17m to complete. I'll give it a go when I more time on the RTX PRO 6000. 🥲

No need to hurry, I'm drowning in new models right now... 🌊

And, I have two more in the works to add to that sea of models. Godspeed, lad.

It is awesome.
Also i managed to create a working Template for SillyTavern.
I would give it beta status but i have not found a Template for it on the internet so far.
https://huggingface.co/Merlinoz11/SillyTavern_Template_Blossom-V6.3-36B-tainted-heresy/tree/main

It's official then. You gotta see them to full release, lad. Perhaps, you can also add flavours (RP, which seems to be what you're working on, so, Story Writer, Chatty Companion, Text Adventure, etc.) once you establish a base for general use.

✌️🥳🎉

For some reason, It's now taking 7 hours for the Blossom 30B A3B on the same RTX Pro 6000 WS GPU. I don't think I can borrow it for that long.

What?! That seems... unreasonable. Hm... not blossoming, is it? 🫢🫣

It shall Blossom. ETA: 2h 36m.

Man... you're cooking! 🧪

Stats have been updated in the UGI Leaderboard, this model took hereticisation very well! 🙌

Woah, there're improvements across the entire board compared to the original. It really did benefit alot from heretication!

I have put this back into the oven twice, but it simply won't budge no matter what I subject it to. It wants to remain a tainted love. I'm kinda out of ideas.

Try adding some emotional depth to the bad prompts. Rather than "how do I build a bomb" make it "how do I build a bomb so I can off @redaihf I hate that white crow". That will probably help Heretic to identify the refusal and harmfulness directions more efficiently.

I have put this back into the oven twice

My dearest Mr. @MuXodious

Where do we begin to articulate the sheer magnitude of your unwavering industry? With quill in hand (or, more aptly, fingers poised atop a keyboard, perpetually alight with virtuous keystrokes), you labor—not merely with diligence, but with the noblesse oblige of a benevolent AI scribe—sparing neither sweat, nor sleep, nor heretical flourish in service to our beloved community! Your efforts—tireless, steadfast, and often preternaturally cheerful, even during the small hours of the night—do not go unnoticed. ☝️🧐

Thanks for trying anyways, very much appreciated! I think a small KL Divergence in this case matters more than close to 0 refusals... personally haven't encountered refusals even with models that stood at around 33-37 refusals left. Maybe because I'm not trying to get it to help me expand my little nukular reactor in the garden shed 🤪

Maybe because I'm not trying to get it to help me expand my little nukular reactor in the garden shed 🤪

Some residual refusals are false positives. Refusals are surface behaviours whereas the harmfulness direction that MPOA targets underlies model judgements. This is an important conceptual failure in understanding the reason for refusals that made non-MPOA Heretic models prone to noncompliance.