Cool!
#1
by kbuettner - opened
Not sure what this one was intended to do, but it performed very well on a private C/C++ code analysis eval, beating out most runs of the base model.
Oh, that's an interesting result!
This is a creative finetune of Qwen3.6 27B for roleplay / writing. In testing currently so there's not a more detailed readme.
There's small snippets of code in the training set, but not a lot, just enough to try and ensure the model generalizes and doesn't catastrophically forget anything so that it still works as a passable assistant. Maybe the reasoning traces present in the data generalized well towards code.