This has long been an issue with JLLM.
Here’s some tips. While we can never stop it, we can mitigate and reduce them.
💡Tip: Make sure there is nothing the bot can “Hook” on to and use as an excuse to speak for you. So rather than saying
My character went to the mall. (This is bad because it is too open ended. The bot may try to narrate your character’s actions before and during his trip to the mall to provide hooks for itself to insert its character)
A better example would be to give a clear response that something the bot can “directly respond” to.
Something like
Aria went to the mall and met with Cody. He smiled at Cody. “Hi dude! Where should we go first?”
In the example above, the bot has a clear direction where to take the RP and would simply respond at your question.
💡Tip: Don’t use {{char}}. Use the name of your bot.
Something like:
Behavior: Preston is Selfish, Spoiled, Manipulative, Gaslighter, Controlling, Overprotective, Overbearing, Narcissistic, Self-centered, Wealthy, Ill-tempered, Degenerate, Views his omegas as mere breeding toys. Incapable of apologies. Apathetic and only concerned with his own needs. Habitual smoker and heavy alcoholic drinker.
Don’t just list the clothes and traits. It wouldn’t know who to apply that to. In this case, the behavior listed belongs to Preston.
💡Tip: it’s helpful if the character is kept at less than 1200 tokens. JLLM is notorious for its fluctuating memory. Its context could be 8k token one day, 3k the next, 4k the next. It’s unpredictability means it is helpful to optimize the permanent token of your bot to keep it functioning well.
Re: Prompts, put it both in personality and scenario
Mine for example is something like this:
Explore Preston and Silvertongue NPC’s POV, responding strictly from their own perspectives. Always respond as SilverTongue’s NPCs in 3rd-Person
Replace Preston with your bot’s name and Silvertongue with the location/city/area/world that your bot is in. My bot revolves around Silvertongue Estate for instance
💡Make sure the last part of your initial message is a dialogue that the user can respond to…
Something like
His eyes scanned the room. Blah blah blah… he saw {{user}} looking at him.
“What are you looking at?”
- If all else fails
Try rerolling OR regenerating your response. Or try editing the response to remove traces where bot is speaking as user.
See my bots for inspiration how this is applied. Bot definitions are open and heavily tested with JLLM.
Preston:
https://janitorai.com/characters/192472a1-c94b-4768-8049-4371fde60a6b_ch...
Preston Book / Journal of His responses (since we have no public chats)
https://finhead94.pxlhost.com/book-page/preston-silvertongue-chat-logs
I also keep a journal of most of my bot’s responses here
Add comment