I’ve noticed Leo has trouble with things written in the margins: my documents frequently have notes or additional info in the margins, some contemporaneous and some added later. Leo either doesn’t transcribe them at all, or seems to transcribe them at quite random points and bury them in the main text, and I then have to work out which marginalia goes where and when Leo returns to the main text.
It would be useful if marginalia could be put all at the end/start, or in a separate box, or some other solution I can’t think of - burying them in the main text is extremely confusing at the moment.
Here’s how marginalia should be handled (in angle brackets with line breaks):
If possible, could you share what’s going on in your transcripts? If this is a widespread issue then we can make it something that we focus on when we scrub (i.e., resolve inconsistencies) in the training data for the next model.
It’s not in every transcript - some of them are fine, but for example this one just ignored the notes in the margin and didn’t attempt to transcribe them.
Thank you Clare! I’ve made a note to look into this when we get around to scrubbing the training data for the next version of the model.
I’m having a similar issue with recently uploaded documents. The response to marginalia feels a bit random to me — sometimes, rarely, it is reproduced in angle brackets like it should be, but more often the text in the marginalia is either just not transcribed at all, or else it is inserted into the main text seemingly at random. This happens within the same document, even when there is no (for me) visible difference between how the marginalia is presented from one page to the next.
I noticed that Leo inconsistently transcribed vertical writing in the margins.

