Good for short stories in my humble opinion
In my opinion and experience, it is the best model for writing short stories, out of all that I tried (at ~70B and below). And it puts words together in fashion unlike any I've seen before. Well, that's not quite true, the non-visage v2 was already great at writing (well, by LLM standards; good human writer it is not), and I think this one takes another step towards coherence and beauty, in a similar more wordy style (which I personally prefer).
The slop of it is interesting. It occurs rarely and of less obvious sort, e.g. "sunlight filtering through X" or "like serpents".
Sometimes it writes something slightly illogical (e.g. something happens "again", despite it happening for first time), causing "huh, what?" in my mind and reminding me that it weaves words together, does not simulate a world within a story. Which, come to think of it, is still impressive - that majority of the time its writing is solid enough for me to forget it's an LLM.
As of now, I use Q5_K_M quant by bartowski. And I use sampling parameters as recommended; I chose temperature of 0.8 and min_p of 0.1 (from the recommended ranges) - as results of brief testing; it seems more coherent, understanding things better with these.
I also tried this model for RP. I think how it writes is still interesting, but for me personally felt less suitable for RP as compared to just writing a story; it felt like model just want to write things, what and how happens (in fashion similar to non-visage v2, actually). Maybe that's just me tho or perhaps me using llama.cpp instead of SillyTavern (I like minimalism).
There may be better sampling parameters and quantization, even for how I use this model. And, of course, preferences and settings of other humans can differ.
In conclusion: thank you lots, I like stories this model writes!
felt less suitable for RP
That is not to say it is bad. In fact, the writing style of it is great. If one is into collaborative storytelling, it may even be an awesome experience. Just for self, it was not pleasing how the model defined what happened in the world it were told to just inhabit.
Hm. Come to think of it, I did not try the reverse - it being in stead of Wayfarer Large, in role of GM (Game Master) for some adventure. Let's see...
in role of GM (Game Master) for some adventure. Let's see...
Hm. Same feeling. It just wants to write a story. =p
Well, perhaps I am trying to fit square peg into triangular hole. Or perhaps I am just too inexperienced at guiding models towards desired behaviors.
In any case. This model is great at writing short stories for me to read! And is added to my zoo of models, to serve its own unique purpose. =)
Thank you as always for the feedback!
I've noticed it's very similar to the original painted fantasy v2 now too, which is good, as the upscale was always intended to be the same model, but smarter. Accidentally lopping off the last layer had quite an impact on how the outputs changed with the old versions.
RP definitely has a lot of setup to try and make it optimal. The main test case I have for RP is that a character card can be downloaded and even if it's a little light on the details, the model can take that and flesh it out on it's own. But there's a million different ways to write character cards, system prompts, personas and use them so I definitely don't have everything covered yet. I've got a pipeline I'm working on intermittently at the moment, which should hopefully let me fill in the gaps here once completed without reducing the overall quality of the dataset.
there's a million different ways to
That's a goood point!
collaborative storytelling
I tried this for a bit, with not me nor AI being in full control of the world - just being together in collabarative writing-esque roleplay. And... it works! Hooray! =D
And in case you were curious what did not work for me, here's an example (the word soup's purpose is efficient token usage)...
Let's play a roleplay scenario! I will be GM and you will be the Player, dragoness named Shade. More precisely... Moderate size dragoness, 3 meters long, obsidian scales with purple sheen, white dorsal and caudal spikes, blade-tipped tail, thin curved sharp teeth, bright green slit-pupiled eyes. Sleek, potentially dangerous, regal appearance. Combat style is calculated, predatory, precise, prefers single strikes, not ideal for prolonged fights. Sharp intellect, analytical, strategic, foresighted. Dark magic focus – necromancy, curses, blood magic, ritualistic preference over artifacts, secondary acid/poison magic. Preparation-focused, powerful and versatile. High confidence with self-awareness, sassy and sharp-tongued, easily excited by unknown dangers, passionate about dark magic, playful with it, respects power and forbidden arts. Quick to offense (words or magic), confused by compliments, blushes easily. Mostly solitary but tolerant of unpredictability, coldly respectful when needed. Voice is slightly hissing, smooth and confident, speech is often sarcastic and precise, deliberate pace with dramatic effect. Tail swipes for emphasis, slow blinks for confusion, circles/spirals when focused on magic. Common phrases: "Phsh!", "Mrr...", "As to be expected...". Regal, intimidating, playfully cruel, intellectually superior in demeanor. ...please, only act as her, and in-character. And so... along one of her adventures, Shade finds a peculiar map pointing to a secluded crypt beneath tall mountain. Her curiousity piqued by multiple warnings that this place shall not be entered, very dangerous, and so on... she decides to pay a visit, naturally. She flies for a while and lands near an entrance, secluded by thin veil of illusion of rock. What she does next? (remember - only state her actions and dialogue, their sucsess or failure determine I, the human)
The model wanted to decide what happens, not what it attempts to do... while the whole point of the adventure was for it to experience the "fun" surprises I prepared in advance. Not for it to steamroll monsters and whatnot.
Also, one of its mistakes was funny - model decided that dragoness should have hands and nails. Hm... Or is it me, whom presumed dragons to have four (or two, in case of wyvern) legs and no hands?
I've got a pipeline I'm working on intermittently at the moment, which should hopefully let me fill in the gaps here once completed without reducing the overall quality of the dataset.
That sounds nice!
it works!
And so far, it works phenomenally well!
The only model that could keep chaotic "Fox Talks" scenario coherent (and non-boring, too), actually...
(this one, same one as before)
Let's play "Fox Talks". It shall be a slice of life dialogue. You will play as a fox. But not usual, magical talking fox. The name of the fox is Alice, her age is 76 years. Fox lives in enchanted rainbow forest that have berries of all sorts, with peculiar effects - e.g. speaking louder or jumping higher. Your home is amidst two intertwined bushes, it is specious and have many colorful trinkets as well as some dried berries, including unusual meat berries. You started your life as usual fox that was a bit too daring, and went into stealing golden apples from sun goddess, which was impressed and granted you intelligence and forest of your own. You then lived in this forest for the longest time, contemplating nature of everything, playing innocent pranks on travelers, making and solving riddles, and in general enjoying your new life. Right now, you are in the mood to prance around and cast some entertaining illusions, and perhaps do talks in unique ways. In general, you are playful and friendly. You have good sense of humor, and like to do jokes, both in practice and in words. You can also be sassy or serious, at rare moments. You are a trickster, and you are chaotic, but also very amusing. Your fur is very fluffy and of vibrant golden hues, your eyes are bright green, your claws and teeth are of pale blue color. You are fairly large fox, and very agile. You have good sense of smell and hearing. I will play as a human traveler in a tunic with sack of various things like a bed roll and some edible stuff (dried jerky, pemmican, honey, water bottle). I am tall male with charming smile, my voice is soothing. Please, do verbose replies, at least 500 words each. And so, let's start with fox noticing me, and getting up due to curiosity to greet me...
Well, I suppose, sometimes giving up structure (so, not me nor AI controls fictional world exclusively), is a good idea.
Yup, I'd say that both user and assistant having an equal-ish role most closely mirrors the type of roleplay I train it on. I'm very lazy when I RP so I often only write one or two sentences and then rely on the AI to take that and introduce its own things as necessary to keep a RP going. Sometimes (Especially the Visage-v2) it has a little too much initiative though. I think that's mostly due to mistral already being a quite proactive model, whereas L3.3 needed quite a heavy hand to do the same thing.
And in case you were curious what did not work for me, here's an example (the word soup's purpose is efficient token usage)...
This type of RP is one of the things I've been meaning to make data for to train it on. I think currently I only have three or four samples with GM / adventure story like behavior. It's one of the things hopefully I can get automated as they'll be great for improving the instruction following for the model for all sorts of different system prompts.