V32

V32 Changelog

• Release includes multiple fine-tunes for different gamemodes and an automatic model picker.
• Base model trained on all gamemodes, fine-tunes for specific gamemodes.
• Standard gamemode model trained from scratch on osu!std data only.
• Trained for 700k steps with batch size 32.
• Added horizontal and vertical flip data augments to object positions.
• Training on new remote dataset on HF with year range 2007-2024.
• Get beatmap descriptors from Lazer user tags.
• Added label smoothing 0.2.
• Changed position coordinates tokenization: Rough position token for 32 pixel grid + position refinement token that allows for 2 pixel quantized coordinates.
• Added sustain tokens for extended sliders, spinners, and hold notes.
• Optimized model training speed and memory usage by unpadding inputs and balancing batch sequence lengths.
• Add randomization data augment to snapping tokens during training to increase snapping robustness.
• Add conditioning for descriptors, year, CS, and global SV.
• Remove distance tokens.
• Add timing model output for adding redlines.
• Set context window size to 16 seconds.
• Changed slider path tokenization.
• Add Timing as an input context for the model.
• Adjusted training LR curve to end at 0.
• Moved conditioning to decoder as tokens instead of encoder input.
• Adjusted DT data augment probability.