Hi all I’m new here , been using LLMS for a while , I’m curious about seeing if any of the established models would have any ideas on a realistic way to train a model in the MOOSE API framework for DCS. I’m aware that some llms are pretty good at coding in base languages now , but most of the ones I try to use for assistance hallucinate crazy days when I try to get them to do any moose-lua coding.
I would’ve thought with an open project like this (all the source files for the framework are available to the public on GitHub).
I attempted to train a smaller mode Qwen coder 2.5 7b quantized down on my 3090 and whilst it learned some of the app it would still drastically fail or hallucinate at certain points and invent methods and classes etc.
I’ve been really curious about AI (and in particular its ability to code) for a while now , and whilst I’m not a great programmer by any stretch of the imagination I was wondering what other established developers thought about this idea , viability , likelyhood of training to succeed etc.
I tried a couple of methods so far , first was creating a dataset form the singular moose.lua file which contains all the methods , classes and parameters etc , second was to use the ingest method by investing the official gethub repo and finally I tried to use the rag method , none of these gave me very good results ,
For anyone who has even a passing interest here is a linknto the frameworls docs .