Fantastic model. Some questions if I may
Greetings! Thank you for this model. I've tested Transgression in many scenarios, in both it's intended purpose ( ͡° ͜ʖ ͡°) and some general roleplay. It is very capable.
I've been waiting to try out new Safeword models, given my experience with Safeword 2.2 was great. I'm not sure if Safeword is meant to be good at this, but it's very capable in general adventure roleplay. sleepdeprived3 stated that 3.0 was meh, so I'm wondering what do you think of 4.0 and 4.1? Currently, there are no exl2 quants for the Safeword 36b version. Are they coming? And lastly, what do you think should I choose, a 24b 6bpw or a 36b 4bpw Safeword, given I have 24gb vram?
Anyways, it's good to see you're working on your Forgotten series. I always recommend them, given a chance. Keep up the good work!
I would personally pick the 4.0 bpw of the 36B, because based on a blind A/B test of 24B and 36B models it was determined the 36B model was "smarter" and better liked overall.
I will start quanting the 36B safeword, not sure why that was overlooked.
Thanks. I'll be sure to try it out. I'm also wondering, what's the difference between Abomination and Safeword series?
Thanks. I'll be sure to try it out. I'm also wondering, what's the difference between Abomination and Safeword series?
COUGH Safeword is pure smut COUGH... usually.
Sleep Deprived would have a better answer for you. I've linked him to this thread.
Transgression is, iirc, a completely new model.
Edit: Or based on a different dataset, anyway.
Safeword is the model that I train. I tend to "overcook" it, knowing that I can always merge it into other things to tone it down, tweak the flavor and add coherence.
Abomination is the standard merge of Safeword, usually done with about 40% Cydonia or any Drummer model. This makes it far more useable and I highly recommend actually using Abomination instead of Safeword unless you want just absolutely unhinged insanity.
Transgression is something new I tried. It's similar to Abomination but even better. Whenever you see a Transgression available in your size, always try this first. It has proven to be coherent and intelligent with low repetition all the way out to the full 32K context and it allows for a slow burn when you want it but it will still go full tilt with the Safeword depravity you've come to love when you are ready to pull that trigger.
I wish there was a 36B of transgression :(
I wish there was a 36B of transgression :(
Me too! You should make one.
Got it, thank you both for clarifying. Guess I'll get both Abomination and Safeword 36b once it's exl2 is complete, to actually compare them side by side.
Funnily though, I've liked how Safeword 2.2 handles various tense scenes like arguments, brawls and fights. Great language, descriptions and creativity all while without being too dark. Looks like I've only scratched the surface, haha. Good to know the difference between all these models. I'll stay tuned and check out for updates and new models from you and ReadyArt!
And yeah, 36b of Transgression would be nice, too. I haven't tested 24b 6bpw vs 36b 4bpw yet, but I trust FrenzyBiscuit's word on that matter.
4.0 exl2 safeword 36B is up, rest of quants incoming next 2-4 hours.
Greetings! It's me again with some new questions. Probably better use this thread rather than create a new one.
I've had a chance to play around with the following models: Abomination 4.0 24b, Abomination 4.1 36b, Safeword 2.2 24b, Safeword 4.1 36b, Transgression 24b.
And somehow, perhaps a bit to my surprise, I found Transgression 24b to be the smartest of them all, even over 36b models. It better understood the implications and meanings, was more creative and yet coherent with that creativity. I'm no expert and did no testing, just compared them in terms of the same card-scenario context. While I like them all, for me Transgression simply is the best among them.
So this is what I wondered about...
I'm a bit confused with settings templates you provide with the models. I've followed each model's readme and used the appropriate template, changing them as I've changed the models. But the templates do not differ that much from one another. There are currently Mistral-V7-Tekken-E, Mistral-V7-Tekken-Extra-Dry, Mistral-V7-Tekken-T and Mistral-V7-Tekken-T4. Are newer settings templates backwards compatible to all your previous model releases and should be used over older ones?
Your new model Gaslight-24B-v1.0 is out. I don't yet understand how training works under the hood, but I'm wondering if it follows the same process as Transgression did? Will all your future models be trained that way?
May not be a question, but a humble request: it would be very nice to have a somewhat neutral model, not the one tending to turn things grim one way or another. I understand that it sounds like Cydonia, the model you base your tunes upon, or perhaps like some other general purpose tune. Yet I feel like your models somehow do the job better for me. Is something planned in that sense?
Thank you. Looking forward to future releases!
If you liked transgression you will like gaslight. According to sleep deprived it's an even better experience.
paging @sleepdeprived3
Transgression required a merge to bring up it's intelligence level. Gaslight is my first attempt at replicating the flavor and intelligence of Transgression without the need for a merge. The techniques I used on Gaslight can be applied to any model which gives us more options now. The settings... I'll be honest, I'm randomly trying different stuff and people always ask for settings so whenever I post a model I post what I happen to be trying lately. If I post it then it usually worked well enough in testing to merit the post. You can use any of my settings with any of my models pretty much with certain exceptions like you can't use Llama on Mistral. If you want a model that is like Transgression or Gaslight but more neutral then a merge with a milder model might be your best bet. There's a free service you can use to make merges. https://huggingface.co/spaces/arcee-ai/mergekit-gui