Duplicates in Irishman
I've adapted the tunesformer code to use the first "patch level" encoder to compute embeddings for the tunes in the irishman dataset; my goal was to look for similar tunes. In the process I found that many common tunes appear multiple times in the dataset in very similar or identical forms. For example, entries
X: 39414
X: 189311
X:73017
X:117072
X:172503
X:19118
are all versions of "Scarborough Fair."
The tunes:
X:148476
X:168990
X:5119
X:87582
X:95481
X:64918
X:115834
X:42643
are all versions of "Mississippi Sawyer".
The tunes:
X:98916
X:5441
X:164662
X:125420
X:29997
X:69129
X:57311
X:66002
X:180704
X:52394
X:151183
X:133508
are all versions of "The Rights of Man."
It's not clear to me whether every tune occurs, say, 5 times, but this is something to watch out for.
Hi Jeremy, apologies for the delayed response. It's normal for the dataset to include similar but not identical tunes; only exact text matches are removed during deduplication. This might result in well-known tunes appearing multiple times.
For identifying similar tunes, you might find the IDs listed here helpful: https://huggingface.co/datasets/sander-wood/irishman/blob/main/variation_ids.json