Panchovix/Wizard-Vicuna-30B-Uncensored-lxctx-PI-16384-LoRA-4bit-32g
And this.
Wow, a blast from the past. But this is a lora and already quantized, I would need the full model, unquantized.
No worries, I can't find it myself. I just want to try out some of the earlier models for story writing purposes :-)
Yeah, more people writing stories please :)
https://huggingface.co/Panchovix/Wizard-Vicuna-30B-Uncensored-lxctx-PI-16384-LoRA-fp16
well, this one should be the full one. it's queued as well. Cheers!
Feel free to report back (or not) on success or failure :)
hehe, when 16k was a long context...
A lot of the good story writing models, under circa 70b like Psyonic Cetacean and Dark forest are all 4096 :-( I've tried loads of YI fine-tunes and Magnum 32b with my mikupad setup and just can't seem to get it to work. I've tried so many different prompt templates and sampler settings with no luck. So I figured I'll try some of the older models with extended context. Thanks so much for doing these.
Well, good luck! My personal favourites are all miqu based still. My prediction is that you will have limited success, but I'd be happy to be shown wrong :)
My experience has been that Miqus have been the best. I had some luck with Gemma but it was very hit and miss. Which Miqu would you recommend? For context I'm making an adult visual novel with a lot of mild bdsm fetish stuff mainly the D side if you know what I mean?(Otk paddles etc) lol
Forgot to say I have a 3090 so can run i2 quants or 70bs
QuartetAnemoi tends to be very good at doing what I tell it to, i.e. its easy to configure, and keeps most of miqu's analytical abilities. Midnight-Miqu is more "creative"-sounding, but also harder to control (it can be fun swapping it temporarily, but I found it not so great for full stories). The various moist variants are... too chaotic, and Nimbus-Miqu is somewhere in between. I'd try quartetanemoi and nimbus-miqu, and see which one works better.
I also have had very interesting experiences with crestf411's models, most recently, L3.1-70B-sunfall-v0.6.1, which, notably, is the first llama-3-based model that didn't immediately disappoint me. Still, quartetanemoi is my safe fallback. crestf411 has some miqu-based models as well, which are astonishingly different in writing than other miqu-based models. And while crestf411's descriptions always make me scared to think of his training material, his models usually are very well behaved (i.e. don't instantly become cheap pulp porn novels).
When you find some other good model, I'd be happy to hear about it, so I can try it myself. I don't know of any specific bdsm-supporting models, but I suspect that "normal" models should work just as fine.
Of course, all miqu stuff is 70b. I usually run iq3_somethings (with some cpu offloading), and that seems... okayish. Not sure how much worse iq2 will be. On the other hand, long ago I ran goliath in non-imatrix Q2_K and had a blast. A very slow blast, but a blast :) I should try it out in a more modern quant version nowadays.
I did try miquliz and was impressed. If the cresfall ones scare you I might have to try one haha
I'm going to try Miqu-Playmaid as well. You've given me a lot to think about thank you very much for that 😌
My experience with crestfalls models is that they are not always making everything into sexual roleplay as some other models, but they can if you want them to, which is usually what I want ("do as I say, do not turn everything into an orgy"). So I suspect whatever dark content they might contain, they will divulge it only appropriately.
Have to look at playmaid. I probably quanted it, but can't remember :)
Oh right, those models. Yeah, I always wanted to try them, but somehow forgot.
Which models mate?
Ah, and I remember I wasn't very impressed with miquliz. But then, after a lot of qwen2 nor llama3, it's probably the best thing since sliced bread. A bit like going from morrowind to oblivion and thinking, oblivion is so dumbed down, and then going back from skyrim to oblivion, thinking, wow, oblivion is a true old school rpg (not sure you can follow, just ignore :)
Which models mate?
The playmaid models. I am downloading as we speak.
You have definitely uploaded a quant.
Ah, and I remember I wasn't very impressed with miquliz. But then, after a lot of qwen2 nor llama3, it's probably the best thing since sliced bread. A bit like going from morrowind to oblivion and thinking, oblivion is so dumbed down, and then going back from skyrim to oblivion, thinking, wow, oblivion is a true old school rpg (not sure you can follow, just ignore :)
I am more a fallout guy myself when it comes to RPG type games. I thought Miquliz was decent but only tested it briefly as it was so slow. I haven't been impressed by any of the Qwen stuff if I'm honest.
But fallout is completely dead. As RPG at least. Fallout 4 killed it. But the main problem is that there is no steady decline in fallout games (F:NV is a bump in the road), so it wouldn't work for my example :) But yeah, you could do the same with f2 => f3 => f4. Anyway, yeah, everybody says qwen2 is so impressive, but compared to miqu, nope.
Still, ot step back a bit, the real problem is that there are no groundbreaking improvements in general, probably because the architecture is too limited and only gradual improvements towards the optimum are possible. Because, let's face it, miqu also kind of sucks. It's just the best of the bunch.
I really enjoyed Fallout 4, I must be a heathen 😂 anyway I tried Quarter Anemoi and Sunfall Miqu and was very impressed. I think now I just need to get my prompts, world info memory and Authors notes sorted, as I don't think they're as good as they need to be. Thanks for all your help man, much appreciated.im going to give Daybreak story writer a go as well, I'll let you know what I think.
Yes, if you liked fallout 4, gtfo, heretic :)
As for QA or sunfall, if you find something that beats those models, be sure to tell me. In the meantime, I wasn't so impressed by playmiqu, neither the 70b nor the 140b. But yes, different models require adjustments (some models need to be stopped form writing too much, some need a lot of encouregement).
BTW, crestfallens storywriter model is... interesting. Maybe not that good, but you can't stop it from writing, and writing...
And lastly, v000... has recently reimplemented the nearswap algorithm used for quartetanemoi and made a series of experimental models (albeit small) that did show some promise:
https://huggingface.co/v000000/L3-15B-EtherealMaid-t0.0001
https://huggingface.co/v000000/L3-15B-MythicalMaid-t0.0001
https://huggingface.co/v000000/DupletBoreas-7B-t0.0001
https://huggingface.co/v000000/TripletBoreas-7B-t0.0001
Oh, thanks for the update, Nah I love all the Fallout games I have played if I am being honest ;-)
I'll have to check these out tonight. Thanks for letting me know :-)
The 15b models are of interest as i could potentially fine tune those a little myself. I just hope I can stretch the context out a bit.
Also, oddly I had some luck with Theia 21b, what is odd is I was using the Vicuna prompt template accidently.
i mostly ignore any prompt templates - those are for chat mode, after all. I found ignoring it (or at best using the base model template, which for llama 3 is just <|begin_of_text|> for example) often gives better results, because you don't drift into the assistant mode as much, but of course, this varies between models.
but mostly, if an llm is good, it will cope with more or less any prompt format. i often just use instructions: narrator: or some variant, and refer to those in my system prompt. within reason (e.g. llama-3 based models often expect at least the BOT token).
and if you fine tune a model, you know where you can get free quants... :)
PS: if it wasn't clear, i am a big fallout fan myself, but mostly fo 1/2 and new vegas. and i must admit bethesda didn't completely botch fo 3.
Ah, in Mikupad you have to select a prompt template, but I'm considering moving back over to Kobold United as it gives you more control over the sampler settings. I am really new to all this though, I must admit. the way its setup for me or should i say how I want it to work is I have scene in mind and based on the memory and charters personalities I have setup I ask it to write the dialogue as that's where I really struggle,
Fallout 3 was what really got me in to Fallout as I was mostly a console gamer at the time it came out, so I have a real soft spot for it. I spent many hours playing that game and fell in love with the series from that.
i have something like this:
Example instructions: this is story text. "this is direct speech." [these are commands to the narrator]
And then sometimes use [use direct speech, narrate their dialogue] or [be verbose, no length limit] in the instructions, or something similar.
Depending on the model, I also remind it in my system prompt to "do not sumarize until explicitly told, write out story fragments and prefer direct speech." or something like that. You probbaly have to experiment and gain experience :)
The problem is that most of these models are made for ERP, more specifically chat, and are not so used to writing multi-character set-ups, but most are capable. Often, I prefer the base model, but most fine tunes are based on the instruct model.
And yeah, FO 3 was a clear let down to typical bethesda storytelling (i.e. dumb), but to my surprise they didn't dumb down the roleplaying as much as I feared. E.g. your skill still somewhat determines whether you hit or not when shooting. If you liked the universe and stories, you might want to try out fallout 2 - it's turn based and isometric, and old ass fuck by now, but it's of course world famous for the way it tells its story without telling you what to do and leaving you complete freedom.
Thanks for the advice man, I was thinking instruct might be better as I'm giving it an instruction as to what I want it to write and that would work better.
I tried those smaller models and I was actually impressed at how well they write given their size. Happy to check out any other recommendations you have ;-)
I will give Fallout 2 a try when I have some free time, I never really get more than half an hour to myself at the moment with me having young kids and working full time.
I don't mind if games are old etc, I still fire up Total Annihilation from time time time. Apologies for the late reply, I got had a bit too much to drink last night and have only just started feeling human again 😂
This looks interesting
mpasila/Mistral-freeLiPPA-12B