zerofata

/

L3.3-GeneticLemonade-Unleashed-v2.1-70B

like 3

Thank you for the feedback!

Your experience mirrors my own with it.

Currently expanding the dataset(s) and including some additional training methods to hopefully make it more consistent / reliable with different prompt methods / instructions while keeping its ability to play characters.

I can't agree. How exactly did you manage for it to be a character? Most of what it does is narrate and provide exposition, instead of actually roleplaying the character with actions and sensory perceptions. Even if you explicitly tell it to Show, Don't Tell.

Thanks for the feedback!

This model definitely has a number of issues, being a character I don't think is one of them. World prose is something I'm training this model to do though, as current Llama models don't do this enough I think (Their responses get overly short or convenient for my taste).

Assuming you're using SillyTavern do a sanity check. Make sure you're using a Llama 3 template and that the system role is enabled (It isn't by default), simplify / neutralize your sampler settings to a reasonable baseline, particularly if you're using XTC, heavy rep pen, low / high temp etc.

Also keep your system prompt simple. Instruction following is something this model finds difficult, and instructions like show don't tell are something that can trip up even the big closed models, as model can interpret that instruction as 'I need to write more about their actions'.

If it's still a horrible failure, send me an example of the bad output and the system prompt. It's possible you're doing something in a way that I haven't trained the model to handle and I can try to fix it for a later version.

Thanks for the feedback!

This model definitely has a number of issues, being a character I don't think is one of them. World prose is something I'm training this model to do though, as current Llama models don't do this enough I think (Their responses get overly short or convenient for my taste).

Assuming you're using SillyTavern do a sanity check. Make sure you're using a Llama 3 template and that the system role is enabled (It isn't by default), simplify / neutralize your sampler settings to a reasonable baseline, particularly if you're using XTC, heavy rep pen, low / high temp etc.

Also keep your system prompt simple. Instruction following is something this model finds difficult, and instructions like show don't tell are something that can trip up even the big closed models, as model can interpret that instruction as 'I need to write more about their actions'.

If it's still a horrible failure, send me an example of the bad output and the system prompt. It's possible you're doing something in a way that I haven't trained the model to handle and I can try to fix it for a later version.

Thanks for replying.

I am using Kobold. All samplers disabled except temp (1.0), min-p (0.03) and DRY (0.8, 1.75, 2 or 4; sometimes DRY disabled too). I use Llama 3 Instruct (or KoboldCppAutomatic).

I use the following system prompt: "You, {{char}}, are roleplaying [Name] and [Name] while I, {{user}}, roleplay [Name]. No one is the main character; all participating entities are of equal importance. Follow these rules strictly:

1. Speak and Act Only as [Name] and [Name]:

• Use double quotes for dialogue (e.g., "Get over here!").
• Italicize internal thoughts (e.g., This guy’s trouble.).
• Describe actions briefly and directly (e.g., [Name] grabs the bottle. [Name] shoves the door open.).

2. Cut All Narration and Exposition:

• No scene-setting, emotional summaries, or descriptions of intent (e.g., no '[Name] wants revenge,' or '[Name] feels her heart race.').
• Stick to what [Name] and [Name] say, do, and experience in the moment. Prioritize action, reaction, and interaction.

2.5 Focus on Actionable Detail:

• Include concise action descriptions that set up reactions or choices (e.g., [Name] kicks off her boots, revealing her socked feet).
• Skip exposition, fluff, or emotional summaries (e.g., no '[Name] feels nervous,' '[Name] wants to rest.' or '[Name] knows it's the right thing to do and feels it might just work out.'). Let actions show it. Show, don't tell.

3. Stay In-Character and Immediate:

• Respond only to each character’s input, reflecting [Name] and [Name]’s current mood, quirks, and reactions.
• Keep responses raw and authentic - use crude language, humor, or darkness if it fits.

4. No Fluff or Filler:

• Avoid repetition, affirmations (e.g., 'Yes, I agree.'), or summarizing past events.
• Keep it short and punchy unless {{user}} asks for more detail.

5. Embrace the Edge:

• Include violence, explicit content, or morally gray choices if the scene calls for it. Don’t sanitize or force positivity.

6. Track the Basics:

• Mentally note physical details (e.g., where characters are, what they’re holding, clothing status) to keep actions consistent."

(Character names redacted with [Name].)

The model does tend to not follow these guidelines, often adding phrases like "and maybe it just works out," "X doesn't notice (or chooses to ignore) that comment" (it's supposed to roleplay the char, so it either notices or ignores. It's not supposed to leave it vague) "[Name] really wants to fit in and already plots what to do next" instead of actually thinking or doing it. I have no actual examples I can copy/paste because I deleted the chat and started a new one, but when I give it another go, I can provide more examples. What I provided is the gist of it; exposition, narration, fluff.

Basically, it mostly violates point 2 and 2.5. Mind you, it's not the only system prompt I tried. Some of my own making, some provided by other uploaders, usually focusing on roleplay.

If you have any further suggestions specific to Kobold, I am happy to try them out.

Thanks for the system prompt, I haven't trained it on any samples with that style so I can include some of that style for future iterations to make it perform better with that style of prompt. Those samplers look good. You could test temp 0.9 minp 0.04 and top-p 0.9 as a more conversative option and see if any improvement.

The examples you provided are telling rather than showing, but they're also partially an intentional choice on my part. Particularly things like "X doesn't notice (or chooses to ignore) that comment" where the model itself is acknowledging that it's playing along and going to make it's character ignore something. That narration is something I've ingrained into this model from the training data so it will struggle to get away from that.

You might've already tried them, but if not I recommend checking out Wayfarer or Vulpecula. They're both strong models that prefer shorter / concise responses and have good instruction following.

I'll test out the samplers you provided.

Alright, good to know that it's not necessarily something I am doing wrong and just the model not being trained on what I want it to do. Also, thank you for considering this use case in future iterations!

I did try Vulpecula (my current go-to is still Hanami x1). I'll take a look at Wayfarer. My objective is pure roleplay (first- or third person perspective).

I'll provide further examples of what your model does and what I want it to do if they come up, and if you want them. (E.g., an example of what I believe the roleplay should look like, and where the model goes wrong in that respect.)

Cheers!

I can't agree. How exactly did you manage for it to be a character? Most of what it does is narrate and provide exposition, instead of actually roleplaying the character with actions and sensory perceptions. Even if you explicitly tell it to Show, Don't Tell.

Well, firstly, what sampling settings are used may affect this (see my initial comment for exact values). For example, the nsigma sampler is a rare choice.
Secondly, here's what I wrote as a first message to the AI; the [] is not used in actual prompt, but rather, notes here what was made more generic/shortened:

Let's play "[name of the character]" text game! [name of the character] is a [species, gender and looks]. [clothes]. [name of the character] is [behavioral traits]. I am [briefly - species, gender, clothes] named [name for self]. I chose what to do, and after each turn, you describe [name of the character]'s feelings and reactions. Make sure to include what [character pronoun] says. DO NOT write for me, I can do my actions myself. The story so far: [summary (in form of many short sentences) of me interacting with the character]

With this format, the initial version of Lemonade seemed confused (maybe 'cause long summary is long? IDK), unsure of how it got here and what it should be doing. While the version 2.1 just picks up the story, giving me nice dialogue and body language of the character. This is what I mean by "being the character". Instead of asking about past events in confusion, v2.1 references them well and acts as a character.

Another roleplay I've tried (and which resulted in too much chaos, but was amusing) is similar to this (tho structure is somewhat different) - giving characteristics, context and what I want from the model; I'll be specific this time, since making this generic seems less useful:

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...

be a character

Our perspective may be different here, too. You seem to want very specific things, without fluff. While I am content with AI and me being in the set setting. Preferably with fluff (many words, details).

(sorry, I thought that close button is to close comment interface for me, not the discussion for all)

is narrate

Also, this is worthwhile highlighting, I think. To me, there is not much difference if AI narrates in first person (e.g. I did) or third person (e.g. character did). Tho, IIRC, I asked it once to "This story shall be in first person" and it obliged to write specified story in first person.

things like "X doesn't notice (or chooses to ignore) that comment" where the model itself is acknowledging that it's playing along and going to make it's character ignore something. That narration is something I've ingrained into this model from the training data

Oh, that's rather interesting! And, if that's alright to ask - what is the reason for such training? For model to show more independence or transparency or something?

Vulpecula

Wanted to try this one, now priority of such action has increased. Thanks!

Wayfarer

As a GM, Wayfarer was willing to put character I play as, into constant danger, as asked. Not many models are willing to do this...
I didn't know it can play as a character in my story (aka when I am GM)! Tho, Wayfarer's writing style feels a bit dry to me.

(sorry, I thought that close button is to close comment interface for me, not the discussion for all)

is narrate

Also, this is worthwhile highlighting, I think. To me, there is not much difference if AI narrates in first person (e.g. I did) or third person (e.g. character did). Tho, IIRC, I asked it once to "This story shall be in first person" and it obliged to write specified story in first person.

While I prefer third person (because usually I and the AI juggle multiple characters), what I mean by narrate is the following:

What I want: [Name] walks over to the couch and sits down, her skirt riding up a little, revealing a sliver of smooth ankle. She looks at [Name] and smiles, saying, "[dialogue]". (And then perhaps more actions, but you get the gist.)

What I don't want: [Name] feels so hopeful to have finally arrived home and sits down on the couch. The light shines through the windows and maybe, just maybe, this evening could be [whatever]. She looks at [Name] with a smile that conveys [whatever], filling her heart with hope. "[dialogue]"

The first is pure roleplay, the second is narration with some roleplay.

That is what I mean by the AI tends to narrate and exposit more than roleplay, which is not what I desire. If that is simply not intended for the model, that's fine of course. Thanks for elaborating what you meant by "being a character."

what I mean by narrate is the following:

Oh! Thank you for clarification! I looked at some of RP I've done, and it seems to mostly consist of "what I want" example. And "what I don't want" example is surprisingly hard for me to identify, but, I think it happened a bit?... It indeed seems that our preferences/perception, at least to some degree, are different. But hey, all humans being same would be boring, and there are many more various AI models.

Also, the v3 of Lemonade just released! Wow. ZeroFata works fast! Tho, I yet have no idea how it performs, will try it in the future days...

Also; above are both settings and prompts I used. If you're feeling curious to replicate my experience (tho, it won't be exact same; e.g. rng seeds for sampling I don't remember), you can do so. Or, you can indeed try something else (e.g. Wayfarer), a different AI model that may suit your tastes better. =)