4.60 star(s) 5 Votes

fierylion

Member
Game Developer
Jan 18, 2019
182
361
Let me clarify again, the game servers were dead for a good 6 hours after its release because it got 10 times the expected demand :(

Should be fixed now, I sold my kidney to rent more servers
 

fierylion

Member
Game Developer
Jan 18, 2019
182
361
I've never used oobabooga; but you could try the on instead and see if it works better?
You need the chat completion, not the regular completition endpoint

 
  • Like
Reactions: fgv2

fierylion

Member
Game Developer
Jan 18, 2019
182
361
..ma bad?
...am sorry....
Is it okay if you can re-update the game exe? The version here (1.0.0) may not work because I changed my server endpoints from mistral.lyonade.net to mistral3.lyonade.net and I don't think its updated there. So basically this version of the game may not work at all xD
 
  • Like
Reactions: Kalvinis and Lerd0

Lerd0

Devoted Member
Uploader
Donor
Jul 29, 2017
10,631
126,441
Is it okay if you can re-update the game exe? The version here (1.0.0) may not work because I changed my server endpoints from mistral.lyonade.net to mistral3.lyonade.net and I don't think its updated there. So basically this version of the game may not work at all xD
..aight done....thread updated...
 
  • Heart
Reactions: fierylion

fierylion

Member
Game Developer
Jan 18, 2019
182
361
..aight done....thread updated...
TYVM!!

Tbh I'm a little sad, this thread was created a bit too early, like an hour after I made the game page public. I was still patching bugs and stuff so a really bad, non-working version of the game was uploaded and everyone here thinks it is a completely broken game even though most of the issues were fixed already
1739407655745.png
 

fierylion

Member
Game Developer
Jan 18, 2019
182
361
Thank you for the feedback! Unfortunately most of the issues were due to this game thread being made too early, literally an hour or so after the game was made public on Itch.io (uploaders are so fast :O). Usually I wait until the game is stable enough before requesting a game thread on f95. All of the errors here were because the original game server (my potato PC) died, likely because it got hit with 11k players on itch.io and god knows how many on f95, I was only expecting 2k or less!!

The creator decided to say this to the AI: game_text should be as long as possible! Obviously, the AI will try its best to generate very long text - to the detriment of the game, because often it'll generate over the limit (set by Max Output Tokens), after which it'll be cut-off mid generation, and the game then fails to parse the output.
Hmm on my test the AI only produce very short paragraphs, often only a few sentences. That's why I told it to produce as much game_Text as it can. I'm working with a relatively small model (mistral nemo gptq). It never ran into the over generate issue on my end. I can't say the same for the other AI models.

The way information is sent to the AI. Example: The AI has no idea about the current/min/max values of each stat, because it's not sent. World information is sent… as a massive unclean json dump, with redundant text (does the AI really need to know the entity id) and strange formatting (why does every " come with a \ before it).
This is a work in progress, I cleaned the JSON and removed unused fields (images and in-game desc are scrubbed for example), thanks for the reminder as I forgot to scrub the entity ids. for the \ its because its stringified json, I'm looking into how to remove these when stringify JSON data.


No place to change samplers, which are parameters used to change how the AI behaves.
I'm using the standard openAI chat API which lacks a lot of parameters, this is the price to pay for supporting as many different AI chat clients as possible (openrouter, koboldcpp, other AI providers and local llm clients...)


Gameplay: Not much to comment here, cause there really isn't any content yet. What I hope the creator do in the future with this project is a curated experience.
Yeah the game description says it right there: I wanted to make an RPG text engine, I didn't put as much time into the 4 starting worlds because they are more like tutorials and examples for others to use to create their own worlds! Have you checked out the world editor yet? My next plan is to create a webservice to host worlds created and shared by others, and make the 3D character model editor modular so it can be swapped with different types of characters and so on.

In general, I made this for my community because I got a lot of requests from my players to create more interactive AI experiences after my last got popular, so that they can make their own infinite adventures.
 

derp752

Member
Apr 22, 2019
127
157
I can't get openrouter working. I've followed the guide but it just loads for ages and comes up with an error. The API is receiving requests but doesn't seem to work :confused:
 

fierylion

Member
Game Developer
Jan 18, 2019
182
361
I can't get openrouter working. I've followed the guide but it just loads for ages and comes up with an error. The API is receiving requests but doesn't seem to work :confused:
what model did you use? Most free models can't produce valid json response
 

derp752

Member
Apr 22, 2019
127
157
mistralai/mistral-7b-instruct:free
Keep getting:

[0d 0h] Failed to parse AI response, retrying...

[0d 0h] Failed to complete action. Please try again.
 

fierylion

Member
Game Developer
Jan 18, 2019
182
361
mistralai/mistral-7b-instruct:free
Keep getting:

[0d 0h] Failed to parse AI response, retrying...

[0d 0h] Failed to complete action. Please try again.
hold on, double check the openrouter URL and also don't spam requests, the free model only allow you maybe 2-3 requests per min I think

EDIT: after some further testing it looks like the mistral 7b fails requests when the world is too complex like Veilwood, but works for the more simple worlds hmm :c
 

Fezz240

Newbie
Oct 20, 2018
52
31
First of all, thank you for taking your time to reply.

Unfortunately most of the issues were due to this game thread being made too early, ...
All of the errors here were because the original game server (my potato PC) died, ...
Very understandable.

Hmm on my test the AI only produce very short paragraphs, [...] It never ran into the over generate issue on my end. I can't say the same for the other AI models.
Checks out; I ran into over generation using a local model (llama 8b finetune).

This is a work in progress, I cleaned the JSON and removed unused fields (images and in-game desc are scrubbed for example), thanks for the reminder as I forgot to scrub the entity ids. for the \ its because its stringified json, I'm looking into how to remove these when stringify JSON data.
Good luck with that! May this and other issues be painless to fix.

I'm using the standard openAI chat API which lacks a lot of parameters, this is the price to pay for supporting as many different AI chat clients as possible (openrouter, koboldcpp, other AI providers and local llm clients...)
Again understandable, though even OpenAI compatible should have stuff like Temperature, no? (I ask as a local small-llm runner; nemo for example seems to like to run colder than others at ~0.7 temp)

Yeah the game description says it right there: I wanted to make an RPG text engine, I didn't put as much time into the 4 starting worlds because they are more like tutorials and examples for others to use to create their own worlds! [...]
In general, I made this for my community because I got a lot of requests from my players to create more interactive AI experiences after my last got popular, so that they can make their own infinite adventures.
My main concern is: Are "games" just going to be "Hey AI, here's the description of the game, go brr"? Because if so, I struggle to see the value proposition here compared to just chatting to an AI normally. Yes, there's nice UI, some stat tracking, but for me this won't be enough to convince me to switch, really.
Though, I might not be your target audience anyways.
Still, good luck with your vision of the game!
 
  • Like
Reactions: fierylion

fierylion

Member
Game Developer
Jan 18, 2019
182
361
Checks out; I ran into over generation using a local model (llama 8b finetune).
Wah! I knew llama models are very smart but they are censored as fuck, do the finetunes work in decensoring them? Last I checked llama 8b wouldn't even write about anything related to the birds and the bees
 

Fezz240

Newbie
Oct 20, 2018
52
31
Wah! I knew llama models are very smart but they are censored as fuck, do the finetunes work in decensoring them? Last I checked llama 8b wouldn't even write about anything related to the birds and the bees
No local model is censored because there will come a finetuner to beat it into submission :sneaky:
Anyways, llama 8b finetunes work surprisingly well as rp models, even becoming the mainstream for its size quite a while (before being dethroned by nemo's)
It's llama 3 specifically and not 3.1 though, the latter's creative brain seems too fried to save
 

fierylion

Member
Game Developer
Jan 18, 2019
182
361
No local model is censored because there will come a finetuner to beat it into submission :sneaky:
Anyways, llama 8b finetunes work surprisingly well as rp models, even becoming the mainstream for its size quite a while (before being dethroned by nemo's)
It's llama 3 specifically and not 3.1 though, the latter's creative brain seems too fried to save
Sorry to bother u but do you have a specific local model you'd recommend that's smart and creative? I'm thinking maybe a quantized nemo fine tune like celeste?
 

Fezz240

Newbie
Oct 20, 2018
52
31
Sorry to bother u but do you have a specific local model you'd recommend that's smart and creative? I'm thinking maybe a quantized nemo fine tune like celeste?
I mean, most models you see people recommend, including celeste, are pretty good, a lot of it comes down to preference, and trying to find the best one will be a fruitless endeavor
Personally now I run either Violet Twilight or Wayfarer-Eris-Noctis, but you'll see people throwing so many others (that are also good) around like AngelSlayer, Slush, Rei, UnslopNemo, NemoMix Unleashed, Mag-mell, Magnum, Lyra-Gutenberg, ...
 

MadKingLear

New Member
May 3, 2022
1
1
Got attacked by slimes, ate them. Ate many of them. Eventually became a humanoid Slime Queen and got slime-pregnant somehow by ingesting so many slimes, decided to input that the resultant slime-baby was humanoid. Fast forward to after turning an abandoned warehouse into a makeshift apartment for new half-slime family. Now am the proud parent of many fast-growing half-slime kids, and am teaching them to read The Hobbit. Blip is such a fast learner.

10/10, would lose will to lewd in favor of wholesome again.
 
  • Haha
Reactions: fierylion
4.60 star(s) 5 Votes