diff --git a/spacy/pipeline/transition_parser.pyx b/spacy/pipeline/transition_parser.pyx
index 5a6b491e0..d46bbb150 100644
--- a/spacy/pipeline/transition_parser.pyx
+++ b/spacy/pipeline/transition_parser.pyx
@@ -281,7 +281,7 @@ cdef class Parser(Pipe):
             # Chop sequences into lengths of this many words, to make the
             # batch uniform length.
             max_moves = int(random.uniform(max_moves // 2, max_moves * 2))
-            states, golds, _ = self._init_gold_batch(
+            states, golds, max_moves = self._init_gold_batch(
                 examples,
                 max_length=max_moves
             )
@@ -304,10 +304,9 @@ cdef class Parser(Pipe):
             # Follow the predicted action
             self.transition_states(states, scores)
             states_golds = [(s, g) for (s, g) in zip(states, golds) if not s.is_final()]
+            n_moves += 1
             if max_moves >= 1 and n_moves >= max_moves:
                 break
-            n_moves += 1
-
         backprop_tok2vec(golds)
         if sgd not in (None, False):
             self.model.finish_update(sgd)
@@ -513,21 +512,19 @@ cdef class Parser(Pipe):
             StateClass state
             Transition action
         all_states = self.moves.init_batch([eg.predicted for eg in examples])
-        states = []
-        golds = []
         to_cut = []
         for state, eg in zip(all_states, examples):
             if self.moves.has_gold(eg) and not state.is_final():
                 gold = self.moves.init_gold(state, eg)
-                if len(eg.x) < max_length:
-                    states.append(state)
-                    golds.append(gold)
-                else:
-                    oracle_actions = self.moves.get_oracle_sequence_from_state(
-                        state.copy(), gold)
-                    to_cut.append((eg, state, gold, oracle_actions))
+                oracle_actions = self.moves.get_oracle_sequence_from_state(
+                    state.copy(), gold)
+                to_cut.append((eg, state, gold, oracle_actions))
+        states = []
+        golds = []
         if not to_cut:
             return states, golds, 0
+        lengths = [len(x[-1]) for x in to_cut]
+        max_length = min(min(lengths), max_length)
         cdef int clas
         for eg, state, gold, oracle_actions in to_cut:
             for i in range(0, len(oracle_actions), max_length):