Update benchmarks/yambda/evaluation/metrics.py

These bugs could substantially distort absolute metric values (especially NDCG), but as far as we can judge it does not change the ranking of the models reported in the paper.

P.S. Thanks to Kirill Khrylchenko for identifying these issues

benchmarks/yambda/evaluation/metrics.py

@@ -47,6 +47,9 @@ class Recall(Metric):
             num_positives = targets.lengths.to(torch.float32)
             num_positives[num_positives == 0] = torch.inf
 
+            # there is a bug: we divide by num_positives instead of max(num_positives, k)
+            # this may slightly affect the absolute metric values, 
+            # but as far as we can judge it does not change the ranking of the models reported in the paper.
             values[k] = target_mask[:, :k].to(torch.float32).sum(dim=-1) / num_positives
 
             values[k] = torch.mean(values[k]).item()
@@ -134,6 +137,12 @@ class NDCG(Metric):
     def __call__(
         self, ranked: Ranked | None, targets: Targets, target_mask: torch.Tensor, ks: Iterable[int]
     ) -> dict[int, float]:
+
+        # there is a bug: we compute (dcg_1 + ... + dcg_n) / (idcg_1 + ... + idcg_n)
+        # instead of (1 / n) * (dcg_1 / idcg_1 + ... + dcg_n / idcg_n)
+        # this may affect the absolute metric values,
+        # but as far as we can judge it does not change the ranking of the models reported in the paper.
+        
         actual_dcg = DCG()(ranked, targets, target_mask, ks)
 
         ideal_target_mask = (