Update.
authorFrançois Fleuret <francois@fleuret.org>
Thu, 4 Jul 2024 11:34:09 +0000 (14:34 +0300)
committerFrançois Fleuret <francois@fleuret.org>
Thu, 4 Jul 2024 11:34:09 +0000 (14:34 +0300)
quizz_machine.py

index f00bf9e..c5da586 100755 (executable)
@@ -328,6 +328,14 @@ class QuizzMachine:
             )
 
             if self.back_accuracy:
+                # If back_accuracy is True, we compute the accuracy on
+                # the backward quizzes not by counting how many time
+                # the real prompt A is equal to the reconstructed
+                # prompt A*, but how many time the answers B* computed
+                # from A* is equal to the correct answer. So we look
+                # for the accuracy of A->B*=B for the forward, but for
+                # the backward we look at B->A*->B*=B instead of B->A*=A
+
                 n_forward = input[:, 0] == self.token_forward
                 nb_total = input[n_forward].size(0)
                 nb_correct = (