Can you force a model into a subgenre? Take the base metal model and fine-tune it — for twelve seconds — on one subgenre’s bands. Do it twice, on tech-death and on black metal, and you get two measurably different models from the same parent.
Two twelve-second edits
The same base model, fine-tuned two ways
Tech-death leans chromatic and leappy; black metal leans tremolo-picked, dense, and stepwise. The split you can hear is the split the real corpora show.
Fine-tune: lr 2e-4, ~80 steps (~12 seconds on an RTX 4090). Curated band whitelists define each subgenre — the whitelist is the instrument, validated by a metalhead in the loop.
The two models diverge on every axis the real genres do
| generated | chromatic | tremolo | notes/beat |
|---|---|---|---|
| base model | 3% | 58% | 3.7 |
| tech-death FT | 15% | 38% | 2.1 |
| black-metal FT | 11% | 59% | 4.1 |
| real tech-death | 13% | 33% | 3.8 |
| real black metal | 8% | 69% | 5.7 |
The verdict
From one base model, two twelve-second fine-tunes produce measurably distinct subgenres that diverge on the same axes the real corpora diverge — tech-death more chromatic, black metal more tremolo-driven and dense. A bonus finding: black metal’s loss is half tech-death’s, because tremolo picking makes it twice as self-predictable.
What transferred, and what didn’t
Light fine-tuning moved the harmonic colour of each subgenre convincingly (tech-death’s chromaticism landed dead-on the real 13–14%), but not the full physical velocity — the fine-tunes never quite reach the real extremes. Style is portable and cheap to install; the honest limit is that it moves partway.