testResults.txt

Aviation Complex:
Test UAR: 0.8012
Test accuracy: 0.7888
Test loss: 0.8873

LA Times:
Test UAR: 0.6378
Test accuracy: 0.4073
Test loss: 1.1546

WSJ:
Test UAR: 0.6866
Test accuracy: 0.6108
Test loss: 1.0870


We originally hypothesized that corpuses of differing formality and style may arbitrarily be assigned a lower coherence score-
however, this was assuming that our models would learn generalizable patterns for sentence ordering and coherence. Instead,
it appears that these models learn a domain specific understanding of sentence ordering, whether by shortcut learning or 
common key factors. Consequently, it seems that scores between the truly coherent and incoherent sets fall in one of two camps-
those it understands and can assign confident scores to, or those it doesn't understand and predicts 0.5 for

Fun other tests:
Test accuracy of WSJ model on the other datasets

WSJ on Reuters
Test UAR: 0.6638
Test accuracy: 0.5842
Test loss: 1.1500

WSJ on LA times and Washington Post
Final selection:
Test UAR: 0.5927
Test accuracy: 0.5347
Test loss: 1.3423

WSJ on Reddit
Test UAR: 0.5210
Test accuracy: 0.4212
Test loss: 1.4488

WSJ on Aviation
Test UAR: 0.5072
Test accuracy: 0.2980
Test loss: 1.4635