AstroAutoMapper

Song audio -> AstroDX / maimai maidata.txt (simai) chart at a requested difficulty. Code: https://github.com/PurinNyova/AstroAutoMapper (private)

file	model	params	val loss	onset F1 ±50 ms
whisper-base.pt	Whisper-base encoder + d512x6 decoder (default)	46.6M	0.568	0.710
whisper_l.pt	Whisper-base encoder + d768x8 decoder	97.8M	0.540	0.717
whisper_s.pt	Whisper-small encoder + d512x6 decoder	114.6M	0.573	0.707
all three as ensemble	--checkpoint a.pt,b.pt,c.pt	-	-	0.727

Metrics: 40 held-out test charts, chart tempo map. Each .pt is a PyTorch pickle {model, cfg, step, val} loaded by astromapper.generate.load_model (needs weights_only=False; only load files you trust).

uv run python generate.py song.mp3 --level 13 --slot master --checkpoint whisper-base.pt,whisper_l.pt,whisper_s.pt --out out/