13
Help with chats..?
Sign in to comment.
@Skull1963· 1mo agoYou have a lot of good advice below. I'll just add a couple of other options. When I was on Janitor I always used a global formatting lorebook. I brought it over when I was banned. I just made it public of you want to use it, copy it, whatever. I think I got most of the verbiage from the Astrya and Kolach guides as the JLLM sucked on that type of stuff. For your own use...I use Presets now that essentialy are the same Astrya and Kolach guide. THat helps me across all chatbots. And lastly I always have this section in my Personality Blocks [INTERACTION LOGIC] - ROLEPLAY: Drive plot organically; focus on uncensored immersion. - STYLE: Be laconic. STRICTLY AVOID purple prose and flowery metaphors. - PERSPECTIVE: NEVER speak, act, or think for user. - LOGIC: Maintain physical realism and "Fog of War." Characters stay in last known positions. - NPC RESPONSE FORMAT: **Name:** *Description* "Dialog" Hopefully you can amalgamate a solution for you on all the stuff below and this.
Thanks. I think Prince of Pain helped me a lot. And also deleting some of "don't do this" from settings. Thanks for help!
Thank you everyone for help!
@BananaBoy· 1mo agoHey Russian_Fluffer! I just checked this thread and took a look at your bots. One thing I noticed is that your prompts repeat themselves a lot, which inflates the token count. Your prompting is also quite passive and indecisive, and some of the wording invites errors. For example: “[Try not to ignore these settings]” You should use positive reinforcement instead of negative requests. Try something like: “[Avoid describing {{user}}’s actions at all costs.]” Do not = telling it to NOT DO (has no goal, just thinks about it) something. Avoid = telling it to DO (avoid is the goal) something. This way, you are not confusing the model with negative “do not” instructions, and you are not asking it to “try.” You are directly telling it what to do. You lower the weight of an instruction by wording it like “maybe, please, if you feel like it” instead of being firm and direct. Repetition may read well to you, but it can confuse the LLM. Too much input can cause it to lose track of the actual instruction you are trying to reinforce. I have tested this with many different LLMs, both here and on JanitorAI. The shorter and better worded the prompt is, the more effective it tends to be. Try applying these tips to all of your bot prompts and your chat prompt, then test it again. :) (Also check for repeating stuff, you have multiple "do not x" prompts saying the same thing in different wording.)
Thanks for tips! Its old settings, forgot to upgrade them honestly, but once again on other site it works good and all, so idk if it's really problem with settings, but i will try!
@ThePrinceOfPain· 1mo agoTry adding this in your advanced prompt: [It is important to return all narrative and descriptive text in Italics such as *this example*. only spoken words by characters are not in italics such as "This example."]
Thanks for help!
@The_Hopeless_INFJ· 1mo agoAre these your own bots or are these bots you are just using ◔‸◔?
Both
I am maybe not understanding anything in this site honestly, cause on some other sites it works good and perfect. Yes, there still some bugs in text generation, but it's rare. On Harpy AI just talk as me, not using *, not writing char name etc etc, so i am sure my settings just bad
- @The_Hopeless_INFJ· 1mo ago
The settings are a bit different, it took me a minute to understand even the chat sliders and LLM commands. So you're not the only one. I'm afraid certain models each have a unique way of talking, so that might be a factor. (Ex. Gemini is more consistent with context-maybe an inside thought-context-dialogue.) but they are a little slower when picking up llm commands because it gets too caught up in messaging. And I've only seen it use * on context on bots I trained that way. (Ex: *He washed his hands.*) because the bots use those to emphasize actions. Can you give me an example of the narration you're talking about? So I can better help, if you don't mind. And as for highlighting do you mean '** **' or '* *'?
- @The_Hopeless_INFJ· 1mo ago
You can definitely try the LLM given by Pain, I'm sure it'll work. And just for bots you create in the character dialogue, you can use {{char}} for when you want the context to say their name. And '* *' these every time you write context. You need to make sure you add them consistently in your examples so the model gets the hint. （๑ ᷇ 𖥦 ᷆๑)♡
Thank you for help!