Model performance
#1
by
skdrx
- opened
From what I can tell, testing by hand, the model is still bad. It can do math somewhat more and sometimes has reasoning success, but I think the dataset had reasoning chains that were too long for this small LM to understand. Also the rchatml format is a bit confusing, I dont think LM studio fully supports it. Regardless, finetuning and chat formats and all that is a complete joke and mess, someone needs to fix that (maybe ill do that later when I have the time). I think im just going to rerun this with some basic instruct format or something, make the model really good (relatively) at instruction following. Any thoughts?
This needs to be tuned to do like basic completion, I unironically think thats the only thing it can pull off