SWG GPT?

[BlazinYou]

Im planning to put together a dataset to train a LLM, the structure isnt set in stone. use unsloth to train either gpt-oss-20b or qwen3-coder-30b. they work on my 24gb 3090.

overall it will be a local model, ran on lm studio or ollama, converse with llm and it can logically generate code or perform actions in SIE and Utinni.

i converted the files of core3 to a `input : output` of `filepath\filename.ext : file contents` im going to train just that to see how it does.

assuming its okay ill need to make api plugin for SIE then MCP server, maybe utinni too?

i have to look into the legality of training their data, but in a perfect world the tres are unpacked and `filepath\filename.ext : file contents` also trained.

training a image or 3d model will have to be a different nerd if its possible or worth it. im already reaching with this.

thoughts?

 

[SolcoGalacticEngineering]

I think it sounds very interesting. I don't know enough to comment on the method or the plan but I hope it works out.

 

[BlazinYou]

ive stayed away from RAG, but everything is pointing me to it and im thinking its the way to go.

list of ignore lines containing this text "remove license for cst/ast", ollama parallel requests, incremental indexing(updates when file is updated) embed files, tree-sitter for CST versions, CLang for AST, put into a chromaDB or something or other blah blah.

found a repo, CodeRAG. im working on adopting that to work seamlessly for Core3/Engine3 (.cpp .h .lua)

then be on track? but then again openai and qwen3 have nice platforms that can be modded.

if i am able to make a dataset and make a lora, gpt-oss-20b is doable for anyone for free on colab im reading. im not sure about qwen3, but that is the stronger fill in the blank coder model.

i need a nerd to bounce ideas off of <3

 

[dsrules]

thoughts?

Are you going to train it on the SWG binary itself? Are you going to train it on the client source code? I'm having a hard to reconciling what the purpose of this is and what problem this would solve. You can't just feed it the Core3/Engine3 source code, it doesn't know what the client is doing. Based on your chart and your description, I think you may have a general misunderstanding of how SWG works. The inclusion of utinni is most baffling, if you train an LLM on Core3 it is NOT going to know what to do with Utinni.

You have some interesting ideas and seem to have a good understanding of the "AI" aspects of this but MAJOR confusion about how SWG works. I would go as far as to say that including Utinni and SIE in this is completely pointless. Hooking SWG is very niche, if your LLM could pump out "useful" Utinni plugins I'd eat my hat.

 

[BlazinYou]

I'm having a hard to reconciling what the purpose of this is and what problem this would solve.

as far as problems im solving i dont know, just creating new ones i guess. that's why im asking for input.

i couldn't write a hello world script to save my life today if i tried. it took me 8 months to figure out how to start up Scurbys devenv in 2010, i was laughed at an told i wouldnt be any use to the project. im youtube university taught on coding, ill never get paid for anything that i would produce, i get that.

i dont know terminology and technical jargon and that is a hindrance for me trying to communicate the thing thats ever evolving in my mind in the terms of whatever it is im trying to make.

the pic i uploaded is outdated at this point.

i need a new tool i guess if i want to do what i want, a structured text dump of iff files (and others) so i can parse into nodes and edges of the knowledge graph( a system thats bidirectional so i can convert the generated "iff files" into the real thing). i wanted to use SIE since its the only tool i know and its already got the structure set up to do everything needed. ive talked to sytner already and api is out of the question for SIE. its the only thing i really wanted tool calls for anyway, to let a llm have the ability to edit/create files and build tres and everything else SIE can do.

i just want to use todays crazy technology(cline + rag) to be able to make the appropriate files if i tell a llm "make a new state like dizzy or stunned, call it stoned, you get it when you eat the new version of bantha butter called Baked goods" "clone the tusken rifle but use t21 stats" "plan to make vehicles have tiered components like ships, engines and armor, complete with radial menu at vehicle repair centers" and it will automatically know what to do.

<3

(ive been using iff as a general extension for all the extensions that can be represented/converted as a text value, i guess)