If some words are indeed "separable" (as suggested by M. Zattera's article), maximizing the F1 score or any metric is futile: the model is not meant to match "separable" words. Fitting the model to data containing "separable" words is probably what creates the wrap-around effect in the MZ slot sequence: slots 7-11 have 6 glyphs in common with slots 0-2, resulting in multiple possibilities of separation: for example sheolkeedy (in f115v) could be, according to Massimiliano Zattera's 12-slot sequence:
sheol keedy
sheo lkeedy
she olkeedy
These 6 words exist in the VM: it is unclear where the space should be inserted. There could be a rule, for example, to choose the first of the three because it maximizes the length of the first word, but we don't know.
Maybe a non-ambiguous slot sequence, allowing a unique re-parsing from a space-less transliteration, would better capture the partial ordering principle at work in the VM.
sheol keedy
sheo lkeedy
she olkeedy
These 6 words exist in the VM: it is unclear where the space should be inserted. There could be a rule, for example, to choose the first of the three because it maximizes the length of the first word, but we don't know.
Maybe a non-ambiguous slot sequence, allowing a unique re-parsing from a space-less transliteration, would better capture the partial ordering principle at work in the VM.