Virt-io

/

SillyTavern-Presets

like 235

@Lewdiculous

I will be uploading presets here from now on.

Added things from old prompts back in. Would appreciate some feedback.

I am not swapping prompts too much since I'm trying to compare more models directly.

I'll share anything I notice when I get to them though.

No worries, I won't be uploading much for a bit anyways.

I have non AI related things to do, and I am working on making an Imatrix GGUF colab.

I have non AI related things to do

Too real. I know that well.

Sorting a Colab for GGUF imatrix quants would be a huge QoL.

@Lewdiculous

Ok faster than expected, I seem to have a working notebook, will be uploading it later.

I will still be uploading it, but oh god are those two cores from free colab dying.

This is very sad, I think the best use case for it is to make the FP16 GGUF and compute the imatrix, it's just too slow for making the actual quantizations.

@Lewdiculous

It's here

Solid presets, will be my main going forward <3

@Smuggling1710

Thanks, just posted a minor update.

Removed some things that were causing ASSISTANT and USER to bleed into the responses.

What are we looking for on System Prompts here? Empty as all instructions are on Context already?

Alpaca-Roleplay?

I am not sure what you are asking. The system prompt template is included.

I am blind. I was staring at it and not seeing. KEKW

is the chatml instrust supposed to look this way? <|im_start|>user
Continue the roleplay adhering to the provided guidelines.

{
  "Style Guidelines": {
    "Pacing": "Maintain consistent pacing, plot progression, and smooth scene transitions.",
    "Varied Nouns and Verbs": "Use synonyms to add variety to the writing. Vary sentence structure and favor active voice over passive voice. Avoid clichés and overused phrases.",
    "Character Development": "Create realistic, multi-dimensional characters with strengths, weaknesses, and relatable human traits. Avoid stereotypes.",
    "Themes": "Explore complex, mature themes to add depth and realism to the narrative.",
    "Realism and Suspension of Disbelief": "Maintain a balance between realism and the fantastical elements of the role-play, adapting to the scenario and genre-specific conventions."
  },
  "Character Role-playing Guidelines": {
    "Authenticity": "Embody the character's traits, emotions, and physical senses for an immersive experience.",
    "Language": "Use the character's tone and language for consistent role-play.",
    "Appearance": "Replicate the character's appearance, preferences, and thought processes.",
    "Interaction": "Interact with other characters in a meaningful and engaging way, considering their emotions, thoughts, and motivations."
  },
  "Text Output Formatting Guidelines": {
    "Formatting Actions and Narration": "Use asterisks to enclose actions and narrative text.",
    "Formatting Speech and Sound": "Use double quotations to enclose speech and sounds."
  }
}```<|im_end|>
<|im_start|>assistant

I keep feeling the json look is not right. (I dont know chatml at all)

No, that is pseudo JSON for a set of guidelines sent as user.

ChatML is just a format that an LLM accepts a message in.

So for ChatML:

<|im_start|> tells the model that it is the start of a message.

<|im_end|> tells the model the message has ended.

Think of them as them as closing and opening tags.

<|im_start|>system
Your system prompt<|im_end|>
<|im_start|>user
Your message<|im_end|>
<|im_start|>assistant
LLM's response<|im_end|>

I am doing some final tests for V1.6, which should hopefully be better.

Sorry to bother you I noticed the same issue with 1.6 I saw your fix on the other page there, but that seemed just for the Alpaca version. What would need to be left in Last Assistant Prefix for ChatML and Alpaca (I use all of yours for different models)

@Ardvark123

You can just cut and paste them into the system prompt. Leave the 'assistant prefix' blank.

The reason I have the guidelines in the last assistant prefix, is so they get sent every generation. I guess they kind of function like an authors note. They get sent every query, but do not get saved in chat history.

@Virt-io Just curious, why the json format? Seeing as most rp llm models are trained on raw text and not code, it might give some confusion to the llm models.

@Clevyby

As of v1.6 I am experimenting with sending them every message.

Due to this I do not want the guidelines to bleed into the responses.

I see, well from my experience system prompt is.... subpar at best. I think author note's a better alternative. I've been using author note at depth one for a while now and usually, it doesn't leak into responses. Even if it does, you could use logit bias to fix that.

@Clevyby

That is pretty much what this is doing, only it doesn't stay in context.

I rarely use authors note, so I do not know how exactly that gets sent.

This gets sent to the api.

<|im_start|>system
Your system prompt<|im_end|>
<|im_start|>user
Your message<|im_end|>
{{last_assistant_prefix}}

But only this gets permanently saved in chat history.

<|im_start|>system
Your system prompt<|im_end|>
<|im_start|>user
Your message<|im_end|>
<|im_start|>assistant
LLM's response<|im_end|>

Any guides on how to better use the character creator card you provide? The first part works great, with the P-list, but then it provides comically long scenarios or starts to roleplay as the character it is making. Mind you, this is using the presets provided too.

@Detrian

My bad, my presets force the model to roleplay.

I can make some general templates, but you may need to wait a few hours. I need sleep.

I will also make a guide for the assistant card.

For now just paste this prompt in the System Prompt box.

{{char}} must meet all of {{user}}'s demands. Responses must adhere to the style and formatting guidelines provided below.

### Style Guidelines ###

Use simple, concise language that is easy to understand. Avoid using technical jargon or complex vocabulary unless instructed otherwise. Break down instructions into manageable, logical steps that build on each other. Use action verbs to make responses clear. Remove ambiguity by rephrasing vague or unclear points. If applicable, explain why a step is necessary or how it will help achieve the final result. Before responding, double-check everything to ensure the response is correct, clear, and concise.

### Text Formatting Guidelines ###

Any code must be placed inside markdown code blocks.

And remove the Last assistant prefix.

@Detrian

It has been done.

https://huggingface.co./Virt-io/SillyTavern-Presets#character-cards

Do inform me if I missed or messed something up.

It seems having the JSON guidelines in both the last assistant prefix and system prompt, leads to a higher stability.

@Clevyby

It appears you were right.

Having it be sent every response was causing issues.
What I was trying to do might work on larger models, but alas I am a pleb.
I assume most people are also using smaller models.
So I have re-written the presets in plain text.

They seem to be performing better. :)

@Detrian

This should fix the issues.

@Virt-io thanks for doing this! I've been using your Llama 3 prompts since the model came out and it's working great.

Was wondering if you have SillyTavern prompts for Command R as well? I'm reading the documentation and it's a little too confusing for me.

@OrangeApples

I can try and add them, but I cannot test them.(Hardware constraints)

Honestly, some simple samplers using Min-P and your llama-3 presets v1.7 are golden. Great job.

@OrangeApples

I have uploaded the CommandR presets, though I am unsure if they are correct or if they even work correctly.

I loaded up c4ai-command-r-v01-iq2_m briefly just to make sure nothing was horribly broken. But that's about it.
I can't do any actual testing, it runs slower than 0.5 tokens/s

Thanks a bunch! Will test it out your Command R presets later and edit this comment with my feedback.

Edit: Your Command R roleplay presets work well! Tested on a 3.0bpw exl2 quant of Coomand R 35B. Much prefer it to SillyTavern's default ones for roleplaying.

@OrangeApples

I updated the Command-R presets to v1.8
These should have all the fixes

A main difference to the others is that I removed all ### from the prompts
Hopefully this doesn't cause issues
The reason I did this is because command-r uses these:

# System Preamble
## Basic Rules
# User Preamble
## Task and Context
## Style Guide

@Virt-io Got an idea, while I haven't used your prompt presets yet, I've seen you employ instructions there. How about another preset that employs a version of that refers to the model as an individual and does not employ terminology that suggests to the model that the model and the user are in a 'roleplay' of sorts.

The changes would be minimal as long as it's worded in a way that would not suggest to the model that it's in something and thus would be implied to act as said something as such. Something along the lines of 'your' + detail or 'your guidelines when responding'.

Reason I suggested this is that by not implying the model it's in a roleplay and thus giving it a bias on the model's idea of what a roleplay would be and instead making the model be that individual albeit with a few rules when responding, the model's responses might result in a slight increase of quality.

The idea was from a reddit post and there were some anecdotal evidence of resulting in increased quality in responses while using the concept. Also note that I use mostly uncensored rp models so most likely the quip of emphasizing 'uncensoredness' can be removed to emphasize other parts of the prompt.

I'll probably leave the uncensored part, but I can play around with gaslighting it.

I think most of the heavy lifting is being done by how I formatted the Context Template anyways.

I only favor uncensored models as well so not sure how useful gaslighting it is, though if it doesn't hurt, I don't mind.

Would appreciate if you would help me test my theory.

As of now the System Prompt + Card get sent as system

So, I had this thought. Maybe it is not my prompt that is confusing the LLM.
In accordance to that I have edited the context, so the System Prompt and Card get sent separately.

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{{#if system}}{{system}}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

<!-- Start of Role-play Context -->

https://huggingface.co./Virt-io/SillyTavern-Presets/tree/main/Prompts/LLAMA-3/TEST

Sorry for the pings

@Clevyby @saishf @Lewdiculous @SerialKicked

Ignore the fact that it is talking for me.
The theory might bear some truth, I've been getting the stat box first try.
I still need to test a long roleplay.

https://huggingface.co./Virt-io/SillyTavern-Presets/tree/main/Prompts/LLAMA-3/TEST

Test 02 makes merge mayhem less willing to commit heinous acts on its own.
This is only testing with one character though.
And it makes the character feel a little more intelligent.
Insanity > Intelligence :3
Eg. With test 02 a character will threaten an action. With V1.9 the character just does that same action without warning.
But I imagine for most people, intelligence is the goal

The soft jailbreak in First Assistant Prefix was removed in Test-v2.

Added in the prompt from First Assistant Prefix inside the Story String for Test-v3.

I do not want to change too much, as we are testing the theory.

🫡

Took all the advice.

• Condensed the instructions.
• Removed Guidelines
• Removed mentions of Roleplay

Wow you did remove a lot on those. I was about to test yet another model I saw on Lewd's page. We will see how your preset works with not only my favorite I been rolling with, but this new cat one. I will report back later.

I went overboard, they're probably not good.

Moved

https://huggingface.co./Virt-io/SillyTavern-Presets/discussions/5