Can models follow a musical instruction?

In short

Only Sol so far marks the crossing and keeps the stems apart in every run.

LilyPond: Espresso Rag hand crossing, by gpt-6.1-sol
GPT-6.1 Sol, run 3. The left hand moves up to the treble staff for bars 9 to 12. I gave it 4 out of 5.$0.05SVGMP3MIDIrag.ly
LilyPond: Espresso Rag hand crossing, by gpt-6-luna
GPT-6 Luna, run 1. The left hand crosses, but its stems point down into the right hand's.$0.00SVGMP3MIDIrag.ly

Finding

The prompt, v1

Write, in LilyPond, 16 bars of café piano music called "Espresso Rag" in which the left hand crosses over the right hand for bars 9 to 12.

In a hand crossing, the left hand reaches over the right, so its notes move up onto the treble staff. The trap is a left hand written high but left on the bass staff. The models did fine and did it in a few ways: move the left hand up, swap the two hands' staves, or write the left hand into the treble staff as its own voice.

Run Checks My rating Cost
Sol 1, 2, 3 6, 6, 6 of 6 4, 3, 4 $0.11
Pareto 1, 2, 3 6, 4, 6 of 6 2, 3, 3 $0.41
Luna 1, 2, 3 4, 4, 4 of 6 2, 2, 2 $0.02

A script reads LilyPond's record of every note (hand, staff, bar, stem) and does the checks. Written after hearing Sol and Pareto though, so Luna is its first fair test.

Verdict

Sol is the only one that gets every check every time, and it sounds the best to me.

n = 9 runs, three models. No test.

The nine runs cost $0.55.

By email

Older, 3 Oct

How jagged is a model?

Newer, 7 Oct

Can Mistral Large 4 make a French press?