AffectScore 1.0 Training Dataset: Valence-Arousal Annotated Music Clips for Affect-Conditioned Music Generation

Curated audio dataset used to fine-tune the AffectScore system -> a LoRA-adapted ACE-Step latent diffusion transformer conditioned on two-layer affect signals for real-time adaptive game music generation. The dataset contains 2,938 short music clips drawn from FreeSound and OpenGameArt, all released under CC0 or CC-BY licenses. Each clip is annotated with valence and arousal coordinates on the Russell circumplex, assigned via a combination of automated MERT-based embedding and manual quadrant verification. Contents: preprocessed_unfiltered/ -- 2,938 audio files (44.1 kHz stereo WAV), normalized and trimmed to training length manifest.json -- full corpus index with per-clip valence, arousal, source URL, and license metadata (2,938 entries) training_set_clean_clap.json -- CLAP-quality-filtered training split (1,571 clips); used as the --manifest argument in all LoRA training runs held_out_set.json -- evaluation split (200 clips, 50 per quadrant, manually balanced); used for MER accuracy and temporal coherence evaluation A three-pass quality filter (keyword exclusion, spectral flatness threshold ≥ 0.05, duration filter) was applied to remove chiptune and 8-bit artefacts before producing training_set_clean_clap.json. The unfiltered preprocessed audio is provided so researchers can apply alternative filtering criteria. Associated code and Colab notebooks: https://github.com/LeeTgk/affectscore