[Deepseek R1] Q1F V1

Download V1 (2025-02-06)

  • Preset
  • Regex: Prettier OOC Formatting (Optional)

    In order to make the HTML dropdowns look a little nicer, you can add this to your theme's CSS. I like it.

    details{
     background-blend-mode: overlay;
    background-size: cover;
        background-color:#0E0C0BDD;
        padding:12px;
        border-radius:4px;
        border:solid 1px #2F2A27;
    display: flex;
    flex-direction: column;
    box-shadow: 0px 5px 5px rgb(10 8 7 / 0.3);
    }
    

Info

This preset is intended to do the following:

  1. Reduce R1's schizo/intentionally disgusting tendencies when in NSFW
  2. Tend towards romantic, sympathetic, and arousing depictions
  3. Works equally well with scenario cards as with single characters
  4. (Optional) Include some light RPG functionality

How To Use

What prefixes in the toggle names do
  • //
    Things that I think are necessary for the preset to accomplish what I intended.
  • +
    Optional niceties. I personally have all of these on at least most of the time, but the preset works without them.
    • + GM Personality +
      I haven't really spent much time getting the personality right, so what's there is very basic. That said it should be easy enough to edit in your own GM personality if you wish.
  • %
    These should be on, but only if your card actually uses the field.
Regex
  • Prettier Formatting
    Aesthetics only, just takes the ((ooc)) command and puts it into collapsible HTML boxes.
Special Commands

You can use these if you want. If you don't feel like using them, then you can either just ignore them entirely or turn off the corresponding + Special Commands + toggle if you want to save the tokens. The square bracket command is part of the main prompt, however, as it provides a lot of functionality and versatility.
If you want to remove the section from the main prompt, R1 has an innate understanding of what commands in square brackets do, but without giving it explicit instructions it can misconstrue their purpose or even start using them itself to change the scene/settings when it's feeling particularly schizo.

Commands

  • [Player Commands]
    If you want to tell R1 to do something right now, format your command in [square brackets].
    • Examples
      • [Skip ahead to 3 weeks later. Give a brief narrative overview of the intervening period]
      • [Tone down some of the zany/comedic elements]
      • [Emulate the style of a Claude Sonnet 3.5 roleplay.]
      • [Until further notice, write in the style of Haruki Murakami]
  • (( Out of Character ))
    When you have something to ask the GameMaster that you want a direct response to, format the query in (( double parentheses )).
    • Examples
      • (( i like this scene, but i'm feeling a bit uncreative. give me a few ideas for directions to take things ))
      • (( how does this character look? what are they wearing right now? ))
      • (( are there any details in the environment that stand out to me? anything suspicious? ))
  • > Story Mode Action
    If you are lazy and want R1 to perform your character, just prefix your prompt with >. These are intended to be like taking an action in an RPG, so failure is a possibility.
    • Examples
      • > attack that character
      • > try to steal something

Recommendations

Sampler Settings

If you use an API that supports temperature, put it between 0.4 to 0.7. No other sampler settings should be needed.

API

I have reservations about pretty much every unofficial API for one reason or another.

  • Official API
    Contrary to my previous thoughts, the official API might be the best. It simply performs better almost every time I try it. It's also cheaper than most other options and supports caching.
    The main downside is that it is currently down basically 75% of the time, and the lack of temperature settings means you need to prompt for more/less creative outputs with Player Commands.
  • OpenRouter
    Basically unusable unless you specify a specific provider you believe to be good. At this point, the ones that I tend to use are Together, Fireworks, and Kluster. You may notice that those are also the most expensive ones. Even if they are the best, they also feel worse than the official API.
    Basically every cheap one, and every free one I have tried are extremely scuffed to some degree. Low quants seems likely, but some of them are so bad that I wonder if they are trying to pass off a distill as the real thing.

My only real API recommendation is to try multiple sources and compare swipes. See which one you like the most. For me, it's the official API. You may feel otherwise.

Why did you make it like this…

  • What the fuck is that formatting? Is this discount W++?
    Yes, I am aware that the prompt is formatted ridiculously. Maybe it's a placebo, but in my testing it seems as though formatting it as if it were pseudo-variables in a script gives R1 the impression that these commands are not things it can ignore/wiggle out of. This also makes the jailbreak more robust.

    R1 seems sensitive to formatting (especially from providers other than Deepseek official API, fwiw). This might not be the ideal formatting, but after some testing I truly believe that the quality of outputs is superior using this formatting over the others I have tested: ASCII, XML, markdown, and org.

  • Why did you make it system prompts??? Haven't you read the readme???
    Sending it as system prompts vs user prompts avoids a problem with OpenRouter. On OpenRouter, user prompts do not get squashed. This combined with strict prompt post-processing forces an assistant message in between every single toggle. Obviously this sucks. If your prompt has 12 toggles, then that's 12 blank assistant messages, each interleaved between your own toggles.
    Using system prompts allows you to squash them, at least, so your pre-chat context is at least one single message as opposed to 12.
    Deepseek official does not have this problem. But in either case, sending the initial instructions as system doesn't seem to have much of an impact.
    Deepseek just doesn't really understand the concept of a message getting sent by the system. It just interprets it as a user message without a name (thus, interpreting it correctly as out of character.)
    If you want to send your prompts as user to adhere to Deepseek's recommendations, you have to:

    1. Use NoAss (i refuse)
    2. Put your entire preset into a single toggle

    I do B anyway for testing because it's easier for me to edit and change things when it's one block of text that I can edit with a real text editor. Which is how I noticed the degradation in performance in the first place. I would finish testing, break it out into multiple toggle because I don't want to force people to edit stuff just to turn off things like the GM personality, and then suddenly my replies would get so much worse.
    At first I thought this was just me running into issues I would have run into eventually, and splitting it out into toggles made no difference. This is how every chat completion preset is structured, I would think to myself. No, it was the act of splitting it that made it worse.

    If you will permit me a little idle speculation: I think that this is basically the only problem that NoAss solves for Deepseek, and that's why it seems to improve outputs for people.
    Trust me, if you don't like the idea of sending it as system (I don't love it either) then just copy every single thing you want into a single user prompt. You don't have to use this preset to do that. Do it with your own, see if you notice a difference. I think you will.

Contact

[email protected]

Contact me to:

  • share logs
  • receive middling technical support
  • complain
  • suggest improvements
  • get music recommendations
  • talk about your waifu
  • talk about your drug addictions
    • n.b. if it is weed do not bother
  • receive life advice from a retard (me)
  • improve my regexes (would appreciate it)
  • etc.

Previous Versions of this Rentry

Edit
Pub: 02 Feb 2025 04:54 UTC
Edit: 09 Feb 2025 10:00 UTC
Views: 8951