lightvector
Question about supervised training process
I’m experimenting with fine-tuning a KataGo model on my own games to capture my personal play style. My plan is to start from a pretrained network roughly at my strength and then run supervised learning over a directory of my human-played SGF files.
When I tried to convert my SGFs into the .npz format using ./katago writetrainingdata, I discovered that it was designed for AI self-play data (it seems to expect MCTS-generated Q-value targets or a rated-games CSV). Pointing it at pure human SGFs triggers assertion failures and other bugs. I haven’t found any branch, flag, or script in the repo that handles “human-only” SGF → .npz conversion. Am I overlooking something, or is the only path forward to patch the existing C++ data-writing code?
In addition, I would also appreciate any insights or experience on human supervised tuning in general (any problems i might run into after having npz training data)!
