The datasets are needed to run anything here but were only mentioned
partway down the card, so the header now names all four pieces a reader
has to collect.
Say the model coverage in a sentence instead of a table
The projections line was bold text rather than a link, so the row mixed
two kinds of entry; all three now point somewhere. Joining the dataset
and acknowledgment lines restores the spaces the Hub was swallowing at
the line breaks.
Reduce the card to the gains a user actually has to set
The variance threshold and the discriminability floor are already baked
into the shipped projections, so listing them alongside the gains invited
readers to look for a place to pass them. One table now covers both
variants; the build-time settings stay in the paper for anyone rebuilding.
Rewrite the model card with a proper header and valid metadata
The front matter had lost its closing delimiter, so the Hub could not read
the card metadata. The card now opens with the project mark and links out
to the paper and the code, and the two hyperparameter tables are folded
away so the usage instructions come first.
Add precomputed projections for five models on three benchmarks
Ships the differential Key and Key+Value projections behind the reported
results, together with the positive and negative decompositions the
baseline and the ablations need. The README lists the validation-selected
gamma, delta_min and gain for every configuration so a projection can be
paired with the settings it was built under.
