@@Thedarkbunnyrabbit I know this is late, but like, it's really not that there's music, it's that your voice doesn't stand out over the music? I'm only saying this though because I like how you organized all this, you've got the chapter markers, and you're so much more thoughtful about this than a lot of creators in this space.
@@Thedarkbunnyrabbit Exactly. People watch this video because they are interested in how to make good cards not in listening to music. It's not that we cant hear you it's that the music keeps taking our attention away due to how loud it is so in a half a hour video it makes my head hurt. Tried to watch it and couldnt, wish there was some way i could download this video and edit just the backbground music out.
4:46 This cracks me up because I was RPing and made a joke about me going back to the "Mystery Machine" to come up with a new plan... well the AI took it literally. When I arrived at the Mystery Machine not only did Daphne and Fred straight-up not exist, but they made a 420 joke about Shaggy on the first paragraph lmao
Thanks for this Bunny. This has been the most helpful resource for me since I started this a month ago... I now have an inkling of what I don't know. Very much appreciated.
I just want to say, I love this video. Very informative and entertaining to listen to. I don't think my character cards were bad before, but they seem much better now. Thank you!
I've recently decided to dive into the AI RP rabbithole by running a local LLM. This is a handy little guide that has some good general information! Audio could be a little better, but it has a lot of information that could put someone into the right tracks towards creating and optimizing characters!
Amen, you're spot on. I've been messing about with ST and locally run models for months now and I usually end up rewriting and reformatting cards I download and I've moved to making my own most of the time. More people need to see this, especially the parts about example messages and italics. Example messages are overrated and whoever came up with italics should be flogged. ;) One thing I'd add as a "do" which doesn't concern character cards per se is that what you write and how you write is also very important. Calling these interactions a chat is incorrect. As you've pointed out, it's more akin to collaborative writing. So if your own writing during the RP is lackluster, it will affect the LLM's output. Final note, I see you're a fellow "excessive negative prompt" user in image generation as well as using "masterpiece" and for that, I salute you.
HA yes, though I admit with V3 now I find those aren't nearly as necessary anymore. In V1 and 2 they were hugely important to avoid lumps of flesh every time you tried to get complex. And thanks! It's a good point. I kind of touched on that in a few points for the card itself, but the better your input in the RP itself the better the output. My ideal goal was to help people make cards that, even when they're super lazy and just want to write one word replies, they can still get good outputs from, though. Because gosh, I've met a lot of people who just don't want to put in effort in the RP side.
Some advise about typing things in the negative prompt. You should not put what you want them to be like. You should put the opposite. So, for example. Putting they are not racist. Is wrong. You should say they are racist. Making that statement a negative since it's in the negative prompt. Meaning the statement is not true. Hope i worded this to be understandable. Thanks for listening to my ted talk.
You're really good at explaining this stuff, I would appreciate any follow up on LLM roleplay on any topic. Personally I am curious about how exactly you use Lorebooks
Did not know about italics being bad. Everything else is all over the internet, but the italics part needs more attention for sure. Good video, thanks!
I rewatch this video about once every two months, because there are no others good videos on the matters (most stuff is in the form various rentries). I once again humbly request you make more, would be very appreciated.
@@Thedarkbunnyrabbit maybe go through creating a card the right and wrong way, recreate the same card using mostly lorebook, show us how you get ai acting proactive? Tbh anything that one can listen in the background would be great, as there is plenty of written sources and there isn't a better way of learning than writing and implementing a card yourself. But for the downtime, listening to you thinking loudly about this stuff would be ideal.
@@Thedarkbunnyrabbit I personally struggle with making bots which keep a consistent way of speaking. Whether I'm on gpt or claude, bots tend to become annoyingly logical and agreeable. They all trend towards a similar voice. Another problem is making a multi character roleplay. Some people do it with group chats, others use a lorebook and a single scenario card. How would you do it?
@@Mbeluba Group chats absolutely need multiple bots. Honestly all group chats struggle a little because it's still only one LLM working, but with multiple cards it can efficiently access different tokens at different times. Also, I could do some like that, sure.
@@Thedarkbunnyrabbit ah see, on ST discord an experiences user advised me to make multiple character roleplays with a single card and to put all character definitions inside a lorebook. That's supposedly cleaner and more efficient token-wise.
Ok, good tips good video, but... no demonstration. It would be really cool to have a step-by-step tutorial explaining why you should do something and how to avoid doing something. For example, bot rules... It was not said exactly where I should place these rules, for example: "Affection rule should be placed in "scenario" and not in "Description".
As obvious as it seems now that I've heard you spell it out, I think I would almost certainly have acted against your advice of making sure to align everything on the character card with what I ultimate want them to do and say, what I expect from them, hope they will deliver for me.
Can you suggest settings and models for a completely custom character? I have a very specific character that I want to recreate. I am being pretty fussy about how she reacts and speaks and things. This worked well in Character AI but in Silly Tavern its really really difficult to get a smooth experience so far. I know your probahbly rolling your eyes when I say I am a complete newbie.
Sure, I need some more info though. What models are you looking at? What's the details of the character? Character AI and SillyTavern are very different beasts, and settings will depend on the model.
This is definitely the go to video for those who wants to make a chara card for Ai chatbots. However, would you be kind to explain more about using italics and how bad it is in general? I want to know more about this.
Using italics for emphasis is fine. Using italics for actions will cause errors for the bot. Bots are trained off of fiction writing - things like novels and short stories and fanfiction. Both Project Gutenberg and Archive Of Our Own are likely used to build its language core. Because of that, it's really hardwired to use italics for either emphasis or thoughts. It has enough training off of internet RP style to recognize using italics for actions, but it will still try to do the former usage even while it does this. It will cause redundant italic tags and coding to start to build up and error and if you try to introduce thoughts or italicized emphasis yourself it'll only make it worse. Also, as I said, italicized action as a format was designed for dialog heavy, action lite RP. It's mostly for chat RP and things like that. You might say _hugs you_ or something to someone in a text/chat conversation instead of 'sending hugs!', for instance. You do this because in the format of communication you're using, you don't need quotation marks - all normal text is presumed to be 'spoken', and you instead need something to differentiate things that aren't spoken. Italics got tangled up into RP because people started RPing in formats like chat settings, where they pretended to be Geralt of Rivia talking on Discord to someone. It inherently was never meant to be in narrative based RP and so it's unwieldy and causes eye strain when you try to apply it to that. Some people try to correct this by re-introducing quotes around spoken words, but that only makes the italics redundant, so all it contributes are its down sides.
Hey, thanks for the video, just a pointer, I had to decrease the playback speed to 0.75 and turned-on Captions to catch what you were saying. Also, the background music is really unnecessary, I know this is a common bad practice that many RU-vidrs follow, but if you don't, then your videos will shine, plus it will save you editing time..
Just released a video diving into a real crime documentary. It took a significant investment of 14 hours. I'm really curious about your feedback, so feel free to share your positive or negative thoughts. Wishing you good health!
can you share some cards you made that we can use as an example template 😔 im trying to learn how to make a good bot but without an example to compare my stuff to this info is hard to take in
@@ThedarkbunnyrabbitI'm in the same place as OP. I have an understanding of the information but I'm unsure about the implementation. Without something to compare with feels I'm stumbling in the dark. Even a quickly made bot would be massively helpful.
It would go in character description, appearance, or a special 'species' category you'd add after their gender. Probably all three spread out appropriately.
Good video! TBH I'm in full frustration about current formats :\ I remember when the condensed description style came up for AID, and it's working great... But this {[Personality("Cheerful" + "Funny" + "Cringy" + ...)]} is just an abomination. Can someone explain to me when it became popular? LLM's called "Natural Language Models" for a reason. And yes, most can understand code formats like arrays and dictionaries. But the example above is none of it... and yes, it will still work, mostly, especially if LLM was trained on this format of bio ( but just OMG why...? ), but this is still just bloating tokens for useless symbols. And here are two examples just for fun. Same character, same bio: first one was formatted as condensed = 360 tokens, second one as JSON (for lolz) = ~580 tokens. That card results are THE SAME and this is just an example of an already optimized bio. + Something like "{{char}} will never talk as {{user}}" is always cute to see. 🤗
W++ was helpful for the 3B and 6B models, like Pygmalion and pre-Turbo 3.5 models, because the models would get confused about what they were reading, and W++ helped separate the character description from the RP or conversation. It also saved tokens compared to using full sentences of natural language. Now, of course, it's just wasting tokens. I don't think there's any still in-use bot that needs W++ to function properly, and some of then it will just mess up. Hah, I've seen some very aggressive token bloats of frustrated card makers repeating "{{char}} will never EVER EVER EVER talk as {{user}}!"
Hey, im trying to make a bot of Lily potter but in an alternative universe that Harry and James are dead and only suvivor is Lily. Naturally she becomes quite mad and desperately seeks revenge while still badly wounded so she is basicly suiciding. As a student we find her in forbidden forrest trying to reach Dumbledore for information and help. And you know the rest but im not sure how much of her previous personality i should write about as she isn't same person anymore. Im hoping bot would realize her change and makes it more insteresting than a mad bot. Also how much about other characters i should write about like Voldemort? Perhaps it would be too long to run well especially in local models i have no idea.
Most bots are going to know Lily Potter, so your tokens will be better focused on writing about this Lily you've made. Emphasize on the ways she's changed, and how she behaves now. You can reference the way she acted in the past if you want the bot to make reference to how she's changed, but primarily your tokens should be focused on Lily as she is now, not how she could have been. You should consider a lorebook for her, so that you can include more information about what all has changed. If you're using local models, try to keep her permanent tokens under 1k and lean heavily on the lorebook to flesh out the details of your AU, in the main card only focus on how Lily behaves and how she looks.
@@Thedarkbunnyrabbit Thank you very much, i was planning to make it usable on ooba etc as well but i guess it is just lost cause. I will lean heavily on lorebook instead as you suggested. Also im thinking to make Snape follow her constantly but he can't face her because how ashamed he feels. Still he defends her against even user if she needs. But ofc tricky part i want bot to not talk about him at all until she is harmed and Spane suddenly appears out of nowhere casting avada around. Do you think it is possible to implement him in such a way? It would be quite fun if possible but never seen such a bot so im guessing it is not sadly..
@@ggoddkkiller1342 You can try to do that, and I do have a character who has a guardian type character, but it doesn't work very well. You'd be better off making two cards and having them interact in a group chat. As I said in the video, the bot gets confused if you ask it to play two distinct characters, and you generally won't like the result. If you want a very basic functionality that _may_ work, you can simply put in a note that says 'if Lily is in danger of death, Snape will appear and defend her/kill her attacker' somewhere in her powers or description section. I've found you really need to invoke it to make it work even then, though. Remember the bot can't be programmed like a game, and is extremely heavily influenced by user input. If you want Snape to appear and rescue her, you want to write in such a way that the most logical thing for it to respond with is 'and so Snape appears' (or similar).
@@Thedarkbunnyrabbit You are really amazing! Alright, i will start from their hideout, how they are happy together even if they have to hide. Then the sad incident and how it changed her. As last forbidden forest while leaving everything else to lorebook. After everything works correctly i will try to implement Snape if it works. Yeah, you never know how it will work out especially with local models but im hopeful. The other day i was trying Queen Elara on MLewd 20B it worked quite correctly. Omg, Leopold was hating her so much, threatening to kill and beat her, constantly insulting her. It was so bad i got mad and literally jumped on him in middle of courtyard lol. So i really wondered what made him hate her so much and read the bot. But there wasn't even hate word in it, simply explaining what is wrong with their relationship, it is such a beautifully written bot. So im planning to adopt it and simply write why Snape can't face her and has to wait for an opportunity to save her and redeem himself, hopefully bot will understand it..
@@Thedarkbunnyrabbit Hey, it is me again :) I wrote the bot it is 1300 tokens 1060 permanent and works kinda well. She is quite mad and badly seeking revenge while acting suspicious and scolding about user etc as planned. But hogwarts plan didn't work so well, i made her still trust Dumbledore and he always convinces her to fight the long battle instead of seeking vengeance and invites her to join order of the Phoenix lol. At least in models i tried so far and even local models like Tiefighter or Beyonder know everything and context constantly jumps around as they pull information from datasheets. When it becomes long term like this you know bot starts generating generic things like they always fought against darkness together or their light was going to endure forever etc. I tried to break it with spesific events but still mostly same can't return to a spesific location and incident entirely. Do you know a trick against this? Perhaps i should make her madder and not trusting anybody. Or drop Hogwarts entirely and make her follow Death eaters alone and we would be an auror also following them.
I gave the most pertinent ones in the video. You can go wild with the biggest lorebook you want, but the entries themselves should be small, and just make sure the tags you use don't trigger too many at once. In Tavern you can turn on the option to let lorebooks trigger other lorebooks. This can be dangerous with a really big lorebook, but it can also help avoid the need to retread info in entries, as relevant info will trigger other entries. Locations, power and worldbuilding specifics, and important NPCs are good options for lorebooks. If you have a long arc with a character either over multiple interactions or one long one it can also be good to have a summary lorebook that tells the bot what it can't remember in a concise manner. This is especially good if key events have happened. Personally I let my characters cook and when they come up with NPCs or background information I like I will add those as lorebook entries on the fly.
I know you're a true crime documentary enthusiast. I invite you to watch the latest real-life crime documentary I shared. Would you like to share your thoughts?
Very useful information but you should be less defensive about the criticism you're getting in the comments. The music is fine, but the volume of your voice compared to it isn't. Wearing headphones shouldn't be a requirement to watch a video, and even doing that doesn't actually solve the issue here. I'm wearing headphones and I can barely understand you (with both system and video volume dragged to the max), when this is not the issue on every other video on CorporateTube. Even if I was patient enough and downloaded the audio from your video, artificially increasing the volume still wouldn't solve anything, because now I would have your loud voice being accompanied by your equally loud "background" music. Practical recommendations from a random internet person: 1) Keep the volume of your voice way above that of the background music. Your background music should be in the background, not in the foreground. 2) Anyone watching your video should be able to follow what is going on without having to max everything or be restricted to headphones to have a chance of understanding you. Medium (see: average) settings should cover your average situation. People should then have the option to increase the volume if they can't hear well or decrease it, if their system is already loud enough. Max settings to the extreme shouldn't be the starting point. 3) If the feedback you're receiving starts forming a particular trend against what you consider to be true, consider the remote possibility that they're right and you're wrong. 4) There's nothing wrong with being wrong. There's plenty of wrong when you deny it. For example, I remember being wrong once - easy 5) The most effective way of being right all the time, and annoy everyone around you, is to always insist on opinions you know to be 100% correct. Being strategic about when to open your mouth is a skill very few people master. Give it a try. 6) When you receive a particular targeted criticism, misdirecting it to something else and pretending that it was the original point all along, just because you feel the need to have a response, any response, is a 100% effective way of having people avoid you like the plague, even outside the internets. Humans can be extremely resilient and they can tolerate many different things, even stinky BO. Stinky attitudes, not so much. One could even suggest that behaviors like these is why aliens haven't made contact with us yet - do better if you want to meet ET in your lifetime. 7) Disregard everything I said, if being corrected by strangers on a topic you're clearly wrong about is some weird preference kids these days have (I should try that in a bot).
I'm not defensive? You can barely hear the background music. The volume is extremely low compared to the voice. Apparently there is some strange set up possible where this isn't true, but even when I max out the volume with headphones the music is still quiet. If anything, you need the headphones to hear the music, not the other way around.
@@Thedarkbunnyrabbit My apologies, I thought I was commenting about the person who made the video titled "Level Up your Tavern Cards: Pro Tips" because if you're claiming you're not defensive about your responses in the comment section, we're clearly taking about a different person and/or universe. I must have left my comment in the wrong video again, or CorporateTube is being silly again today. Speaking about that other person, not you, you can indeed barely hear the background music in that video, which also happens to be true about the narration. You can barely hear either of them because they happen to be almost on the same level, which on its own would be bad, but on top of it, the volume is so low it is barely perceivable to your average bat. Again, I would like to thank that person, not you, for the helpful tips, but that person, not you, should probably do something about the audio.
@@arispaparis Sorry, but being frank is not being defensive. The volume of the video is not excessively low, and even if it were, that is nothing to do with the complaint of the background music being too loud for the voice. If you find that it is too quiet, it is an issue with your set up, not mine.
@@Thedarkbunnyrabbit sorry but being delusional is not being frank. The amount of feedback that you've gotten under your video about the volume being too low, is not due to a conspiracy orchestrated by the easter bunny just to mess with you. If it was a problem with all those random people, not you, you'd be seeing these complaints in other videos too. Unless, maybe it's not the easter bunny conspiring against you, it is CorporateTube! It found all the people in the platform with audio issues, and decided to point them to watch your video, which just so happens to have pristine audio levels, truly. That, or you're being overly defensive and/or delusional. I don't know it could go either way. I just can't tell Also, speaking of you being defensive and/or delusional and trying to misdirect people's criticism to something that your fragile psyche can process, I've already pointed out that the volume of the video is excessively low AND that the background music is too loud for the voice. I am perfectly aware that those two things have nothing to do with each other, and yet your video happens to be plagued with both! So, indeed a video can have perfectly fine volume levels and yet the background music can be too loud for the narration. It can also be true that a video can be too quiet, without the background music being too loud for the narration. Would you believe it that it can also be possible in this universe that we both inhabit (although I'm not sure about this one) for both of these things to hold true, even if they're unrelated to each other? Owning a dog has nothing to do with being called Jack, and yet there are Jacks in this world that happen to own dogs. I know, shocking. In conclusion, a) your background music is too loud for your narration b) the overall audio volume in this video is too low c) both of those things can happen at the same time d) all those people who have independently come over to watch your video and complained about the audio are doing it for the reasons a, b and c e) the claim that all those people are having issues with their audio setup and not you, is laughable f) anyone making that claim is being overly defensive e) anyone claiming that they're not being overly defensive is being delusional