If you can’t delete a concept, can you add one? Run deletion in reverse: take a classical model, which has no blues in it, and edit blue notes into its training data. It works — and the way it works is the honest, slightly deflating part.
The edit, at two doses
Over the major and dominant harmony in MAESTRO, lower a melodic third or seventh to a blue note, leaving the chord’s own third underneath so the harmony stays major and the note grinds against it — the authentic inflection.
Classical, with a blue note injected
These are ear-check clips of the edit itself. The model trained on the dense dose then played blue thirds about five times more often than control.
Generated blue-third rate, 190 pieces per model: control 0.0059; clean 0.0083 (1.4×, z=1.0); loose 0.0286 (4.8×, z=5.9). Real classical sits at 0.0312.
Quantity beat quality
Two arms. A sparse, carefully-chosen 0.067% of notes — the blue notes placed exactly where a tasteful player would — did nothing the model would keep. A dense, rougher 2% lifted blue-note production roughly five-fold. And the honest footnote: an earlier small sample suggested the clean version worked; more data washed that out as noise.
The verdict
Put the two directions together and you get the actual claim: these models cannot be made to un-learn a concept their grammar entails, but they readily learn a new stylistic one you feed them enough of. Easy to teach, impossible to un-teach by absence — and quantity beats quality.
Caveat we caught
The clean-dose “win” at n=40 was a mirage that vanished at n=190. It’s here as a reminder to distrust small generation samples — a lesson that shaped every measure after it.