greedy_until() fix (#344)

Co-authored-by: Nathan Habib <30601243+NathanHB@users.noreply.github.com>
Co-authored-by: Clémentine Fourrier <22726840+clefourrier@users.noreply.github.com>

src/lighteval/models/nanotron_model.py

@@ -1133,7 +1133,7 @@ def greedy_until(
             else:
                 # Longest context in the current split is the first item (since we sort reversed)
                 context_enc = dataset[0][1].tokenized_context
-                max_gen = max(item[1].generation_size for item in dataset)
+                max_gen = max(item.generation_size for item in dataset)
                 max_input_length = min(len(context_enc) + max_gen, self.max_length)
 
             batch_size = self._get_batch_size(