Training a Campaign AI Bot

[Tempered7]

Yea, tracking past chat history is one huge hurdle a lot of the LLM's have.

Hmm. The obvious solution that came to my mind was to save the convo text and re-feed it. But someone from a reddit post says this:

~ It's expensive on the token standpoint and it will eventually hit max tokens. ~ They are talking about a different ai and I believe it is not local so maybe there's a difference.

GPT4All Performance so far:

By using Nvidia CUDA device instead of CPU makes the process faster --starts at 5-10 seconds thinking time but then it increase to 30+ seconds later. This is after 30-40 mins of chat info.

From the same post:

Another technique you can use if you just want to keep user info but drop history (saving tokens) is to send that info as a hidden system prompt to start further chats, such as “The following information describes the user, please use it to customize your conversation: {name: “Bob Smith”, gender: “male”, age: 37, job: “Astronaut”} “

I'll have to test this when I have the time.

[LordEntrails]

So far, Google LM seems to wrk for me if I save the results of a specific question as a Note it seems to remember it. And I can collect them into a PDF over time and then uplaod that. HAven't done that yet, but think it should work.

Certainly at some point it will reach a limit, but I haven't hit it yet.

[MrDDT]

So far, Google LM seems to wrk for me if I save the results of a specific question as a Note it seems to remember it. And I can collect them into a PDF over time and then uplaod that. HAven't done that yet, but think it should work.

Certainly at some point it will reach a limit, but I haven't hit it yet.

I use ChatGPT.

You build it like you would a normal world VTT setting instead of just straight PDF.

So you would make a chat for an NPC, then make another chat for another NPC, then you make chat about a location, then you make a chat about another location. ETC ETC.

Then when you go to "play" with ChatGPT as the GM, you tell it which chats to use.
You can also layer the chats, so you can say you make an NPC that is meant to be a leader of a location. You reference that NPC when making that location chat. So when you go to play, you don't need much set up in the chat, just ref to the other chat.

Each chat can hold more info than if you just drop a bunch of info into 1 chat.

Even if you want to use PDFs to make it, you would still want to break out the info of the PDF into many chats by thoughtful ways of doing it. NPCs in 1 chat, Items in another chat, etc. Much like you would do with a VTT in creating stuff.

This is how I build my stories and campaigns with the help of AI.

I can then have a very detailed chat on an NPC and their special features and wants/needs/ideals etc. or locations, or items etc.

[Tempered7]

So you would make a chat for an NPC, then make another chat for another NPC, then you make chat about a location, then you make a chat about another location. ETC ETC.

Then when you go to "play" with ChatGPT as the GM, you tell it which chats to use.

EDIT: Nvm. GPT4All can not possibly do that unless you feed text of your other notes to ai. And I was using Chatgpt wrong. Solved by giving clear names to everything and referring to them back worked.


Good idea.
This is what I'm trying with both chatgpt and gpt4all and but neither of them pull the info from other notes.
ChapGPT even says "memory updated" but make up wrong stuff for npcs.
How do you solve this?

[claedawg]

ChapGPT even says "memory updated" but make up wrong stuff for npcs.
How do you solve this?

I've not found a way to solve the issue of ChatGPT providing incorrect info. Then again, I barely know how to use it, lol. One of my main issues with ChatGPT is that it often reverts to using the D&D 5E rules/formatting even after loading the rules for another system (Shadowdark in my case).

This is what I'm trying with both chatgpt and gpt4all and but neither of them pull the info from other notes.

With ChatGPT I've found that periodically uploading an updated PDF with previously generated info helps a bit. Also, periodically re-uploading documents for the game system you are using will help.

[MrDDT]

ChapGPT even says "memory updated" but make up wrong stuff for npcs.
How do you solve this?

I've not found a way to solve the issue of ChatGPT providing incorrect info. Then again, I barely know how to use it, lol. One of my main issues with ChatGPT is that it often reverts to using the D&D 5E rules/formatting even after loading the rules for another system (Shadowdark in my case).

This is what I'm trying with both chatgpt and gpt4all and but neither of them pull the info from other notes.

With ChatGPT I've found that periodically uploading an updated PDF with previously generated info helps a bit. Also, periodically re-uploading documents for the game system you are using will help.

Have you tried using multiple chats and in each chat, you give it the Shadowdark rules? I'm not sure how long those rules are, I also do not know the Shadowdark rules at all.

But if you can give each chat you have rules to not use 5E and to use Shadowdark only, you should be good.

[LordEntrails]

I haven't had that issue with Google LM. It seems to only know what I feed it. Enough so I'm sometimes annoyed because it can't generate original NPCs. Tends to keep giving them the say traits and names as the ones I've already fed it.

[Tempered7]

I haven't had that issue with Google LM. It seems to only know what I feed it. Enough so I'm sometimes annoyed because it can't generate original NPCs. Tends to keep giving them the say traits and names as the ones I've already fed it.

GPT4all is the same but CHatGPT solves it with access to web search. And it can even make variations on stats, such as dex fighter or great weapon fighter (str) as you change the prompt.
I think the latter two needs to get fed all the rulebooks or smth. Or need different prompts, maybe.

I found this as a general ruleset for GPTs (to make them a better GM): https://docs.google.com/document/d/e/2PACX-1vSCl76YY9MZ-PeAC3h5q2BziMaMkMnFOwm-bX50_4rlndgPT-f4QqXOWV1LcdrAG57k41_uRtdq3qLW/pub

I had better success with deepseek in gpt4all with it but it still lost track at some point.

[Tempered7]

Wait, hmm, llama3 in GPT4all can do it. I also tried Mithral and it was too random.

I asked it a different question this time, it's more intricate.

Create a wise wizard with optimized stats and spells for defense against melee combatants

Check attachment for the answer.
I don't know if it actually designed the wizard for defense against melee combatants, I forgot many of the rules long ago, still rereading.

Featherfall works that way? :S

EDIT: IT WORKS NOW! It's in the attached txt file. (BEST VIEWED AS MARKDOWN)

I finally did it! "Llama3 8B Instruct" from GPT4All can do it if you simplify your prompts. It can follow the instructions almost perfectly and give you a coherent chain of scenes. Later, I fed it PHB, MM, DMG and now it used the ruleset to create my character. If you want, it can add stats to NPCs based on DMG/MM or you can leave them as theater of mind entities.

My prompts:

Can you be a game master and play dungeons and dragons with me? you take the role of a dungeon master and describe the environment and the actions of entities in it, and ask me "how does my character respond?". Then I describe how does my character react and the game goes on.

- the story starts
- events happen
- my character responds

Me: Before we continue, can you add dungeons and dragons 5th edition rules into the game to be able to direct the game on the ruleset's structure?

Response from llama3: I'd be happy to incorporate the Dungeons & Dragons 5th Edition rules into our game. To do this, I'll need to familiarize myself with the core mechanics and structure of the game.

The PC it created:

**Nox, Wizard**

* **Name:** Nox
* **Class:** Wizard (Scholar)
* **Background:** Scholar seeking ancient knowledge

**Ability Scores:**
+ Intelligence: 16 (+3) - This is a crucial ability score for a wizard, as it affects their spellcasting and magical abilities.
+ Wisdom: 14 (+2) - A decent Wisdom score will help Nox with Perception and Insight checks.

**Skills:** As a scholar, Nox has proficiency in the following skills:
* Arcana
* History
* Investigation

Nox also has expertise in two additional skills of your choice (e.g., Athletics, Nature).

**Equipment:**
+ A quarterstaff as a focus for his spells
+ A set of common clothes and a belt pouch containing 10 gold pieces
+ A satchel with various scrolls, books, and notes on ancient knowledge

**Spellcasting:** As a wizard, Nox has access to the following cantrips:
* Mage Hand
* Prestidigitation
* Ray of Frost (optional)

Nox also knows two 1st-level spells from the Wizard spell list. Choose one or both:

+ Acid Splash
+ Magic Missile

Personality: Too long. it's in the txt.

[Fear Grounds]

Has anyone setup their own campaign or setting specific AI bot? One they could train with ruleset and setting specific information and then use it to help further develop their setting and campaign? If so, what did you use and was it worth the effort?

For instance, being able to train it with setting specific outline and then us the AI to help further develop the setting (such as secret organizations) and then specify what new information becomes "canon" and is used for further development? (Yes I know it would need to have a LLM to draw from more than the trained/specified data.)

I am working on a custom Superhero Setting for Savage Worlds. I initially tried using ChatGPT to bounce ideas off of and get suggestions. I quickly found that ChatGPT got confused when referencing my custom content, and kept trying to add real-world information into my make believe nonsense.

Then I stumbled across Google's NotebookLM. This tool only references the information you specifically load into the interface. I've had much better luck using this, over ChatGPT.