diff --git a/spacy/gold.pyx b/spacy/gold.pyx
index ed66390e4..d344473bf 100644
--- a/spacy/gold.pyx
+++ b/spacy/gold.pyx
@@ -148,7 +148,7 @@ class GoldCorpus(object):
         self.train_path = util.ensure_path(train_path)
         self.dev_path = util.ensure_path(dev_path)
         self.train_locs = self.walk_corpus(self.train_path)
-        self.dev_locs = self.walk_corpus(self.train_path)
+        self.dev_locs = self.walk_corpus(self.dev_path)
 
     @property
     def train_tuples(self):
@@ -173,7 +173,7 @@ class GoldCorpus(object):
             random.shuffle(self.train_locs)
         gold_docs = self.iter_gold_docs(nlp, self.train_tuples)
         if shuffle:
-            gold_docs = util.itershuffle(gold_docs, bufsize=shuffle*5000)
+            gold_docs = util.itershuffle(gold_docs, bufsize=shuffle*1000)
         yield from gold_docs
 
     def dev_docs(self, nlp):
diff --git a/spacy/language.py b/spacy/language.py
index 12964784c..0f6daec70 100644
--- a/spacy/language.py
+++ b/spacy/language.py
@@ -182,8 +182,8 @@ class Language(object):
         for proc in self.pipeline[1:]:
             grads = {}
             tokvecses, bp_tokvecses = tok2vec.model.begin_update(feats, drop=drop)
-            d_tokvecses = proc.update((docs, tokvecses), golds, sgd=sgd, drop=drop)
-            bp_tokvecses(d_tokvecses, sgd=sgd)
+            d_tokvecses = proc.update((docs, tokvecses), golds, sgd=get_grads, drop=drop)
+            bp_tokvecses(d_tokvecses, sgd=get_grads)
             if sgd is not None:
                 for key, (W, dW) in grads.items():
                     # TODO: Unhack this when thinc improves
@@ -228,6 +228,7 @@ class Language(object):
         scorer = Scorer()
         for doc, gold in zip(self.pipe(docs), golds):
             scorer.score(doc, gold)
+            doc.tensor = None
         return scorer
 
     @contextmanager