Potential errors in the Train_textual - how would you like to address them?

#2
by Ryebrye - opened

I'm working on generating synthetic thought traces for these examples to use this as training for a reasoning model, and have found some that seem to have the incorrect answer in the Train_textual:

For instance row 36 https://huggingface.co/datasets/Sylence/SSMR-Bench/viewer/default/Train_textual?row=36

asks what the time signature for this piece is:

L:1/8\nQ:1/8=180\nK:C\n| e2 g2 ga | g3 e c2 | e2 g2 g/f/a | a3 g e2 |\n

The answer marked as the Correct Answer is 4/4 but that's clearly not right.

Measure 1 (e2 g2 ga): Two eighths (quarter) + two eighths (quarter) + one eighth + one eighth = 6 eighth notes.
Measure 2 (g3 e c2): Three eighths (dotted quarter) + one eighth + two eighths (quarter) = 6 eighth notes.
Measure 3 (e2 g2 g/f/a): Two eighths (quarter) + two eighths (quarter) + two sixteenths (g/f/ equals one eighth) + one eighth = 6 eighth notes.
Measure 4 (a3 g e2): Three eighths (dotted quarter) + one eighth + two eighths (quarter) = 6 eighth notes.

You could make an argument for 3/4 or 6/8 - but based on the strong beat patterns, it seems like it should be called a 3/4?

And on https://huggingface.co/datasets/Sylence/SSMR-Bench/viewer/default/Train_textual?row=151

It's asking where to place bar lines:

L:1/8\nM:2/4\nK:E\nB,2 G,>E, [M3/4] F,4 z G, [M2/4] G,F, E,F, F,2 E,>B,,\n

the answer marked as correct is:

L:1/8\nM:2/4\nK:E\n| B,2 G,>E, | [M3/4] F,4 | z G, [M2/4] G,F, | E,F, F,2 | E,>B,, |

but if you look at that, the [M3/4] F,4 has only 4 8th notes in it, and then the remainder of the song is incorrect as well.

The correct answer should be:

L:1/8
M:2/4
K:E
| B,2 G,>E, [M3/4] | F,4 z G, | [M2/4] G,F, E,F, | F,2 E,>B,, |

Measure 1 (2/4): B,2 (2) + G,>E, (2) = 4 eighth notes.
Measure 2 (3/4): F,4 (4) + z (1) + G, (1) = 6 eighth notes.
Measure 3 (2/4): G,F, (2) + E,F, (2) = 4 eighth notes.
Measure 4 (2/4): F,2 (2) + E,>B,, (2) = 4 eighth notes.


Would you like me to keep track of these and open a pull request to address them? I will also do a sweep over the test set once I've completed this.

Sign up or log in to comment