Empirical Study of Pop and Jazz Mix Ratios for Genre-Adaptive Chord Generation
Researchers investigate catastrophic forgetting in music generation by fine-tuning a 25M-parameter Music Transformer from pop to jazz, empirically measuring how much source-domain data must be retained to preserve original performance while acquiring new genre capabilities. The work addresses a fundamental transfer-learning challenge that extends beyond music to any domain-adaptive model, offering practical guidance on data retention ratios during fine-tuning across constrained target corpora.
Modelwire context
ExplainerThe study quantifies a specific trade-off: how much source-domain data (pop music) must be mixed into fine-tuning to prevent performance collapse on the original task while learning a new one (jazz). This moves beyond 'catastrophic forgetting exists' to 'here are the actual retention ratios that work for music models at this scale.'
This directly extends the domain adaptation bottleneck flagged in the arXiv paper from May 6 on reordering data for domain shift. Both papers treat the core problem as data-centric rather than model-centric: how to structure training data when distributions diverge. The music work is narrower (pop-to-jazz, single model), but the retention-ratio finding complements the variance-reduction approach in ORDERED by offering practitioners a simpler heuristic for constrained scenarios. Neither paper solves the underlying tension, but together they suggest the field is converging on empirical measurement of domain gaps rather than hoping architectural changes alone bridge them.
If the reported retention ratios (the specific pop/jazz mix percentages) hold when applied to other genre pairs (classical-to-electronic, country-to-hip-hop) or other music models beyond the 25M-parameter Transformer, that confirms the finding generalizes. If they don't, the result is likely tuned to this specific architecture and dataset pair, limiting practical utility.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsMusic Transformer · arXiv
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.