diff --git a/spacy/cli/init_model.py b/spacy/cli/init_model.py
index 87c3033ad..bad63209e 100644
--- a/spacy/cli/init_model.py
+++ b/spacy/cli/init_model.py
@@ -90,12 +90,13 @@ def create_model(lang, probs, oov_prob, clusters, vectors_data, vector_keys, pru
             lexeme.cluster = 0
         lex_added += 1
     nlp.vocab.cfg.update({'oov_prob': oov_prob})
-    for word in vector_keys:
-        if word not in nlp.vocab:
-            lexeme = nlp.vocab[word]
-            lexeme.is_oov = False
-            lex_added += 1
-    if len(vectors_data):
+    if vector_keys is not None:
+        for word in vector_keys:
+            if word not in nlp.vocab:
+                lexeme = nlp.vocab[word]
+                lexeme.is_oov = False
+                lex_added += 1
+    if vectors_data:
         nlp.vocab.vectors = Vectors(data=vectors_data, keys=vector_keys)
     if prune_vectors >= 1:
         nlp.vocab.prune_vectors(prune_vectors)
diff --git a/spacy/cli/train.py b/spacy/cli/train.py
index 4cad2cae1..a9c332fe3 100644
--- a/spacy/cli/train.py
+++ b/spacy/cli/train.py
@@ -7,6 +7,7 @@ import tqdm
 from thinc.neural._classes.model import Model
 from timeit import default_timer as timer
 import json
+import shutil
 
 from ._messages import Messages
 from ..attrs import PROB, IS_OOV, CLUSTER, LANG
@@ -186,14 +187,15 @@ def train(lang, output_dir, train_data, dev_data, n_iter=30, n_sents=0,
         with nlp.use_params(optimizer.averages):
             final_model_path = output_path / 'model-final'
             nlp.to_disk(final_model_path)
-        components = []
-        if not no_parser:
-            components.append('parser')
-        if not no_tagger:
-            components.append('tagger')
-        if not no_entities:
-            components.append('ner')
-        _collate_best_model(meta, output_path, components)
+    components = []
+    if not no_parser:
+        components.append('parser')
+    if not no_tagger:
+        components.append('tagger')
+    if not no_entities:
+        components.append('ner')
+    _collate_best_model(meta, output_path, components)
+
 
 def _collate_best_model(meta, output_path, components):
     bests = {}
@@ -202,8 +204,8 @@ def _collate_best_model(meta, output_path, components):
     best_dest = output_path / 'model-best'
     shutil.copytree(output_path / 'model-final', best_dest)
     for component, best_component_src in bests.items():
-        shutil.rmtree(best_dir / component)
-        shutil.copytree(best_component_src, best_dest / component)
+        shutil.rmtree(best_dest / component)
+        shutil.copytree(best_component_src / component, best_dest / component)
         with (best_component_src / 'accuracy.json').open() as file_:
             accs = json.load(file_)
         for metric in _get_metrics(component):
diff --git a/spacy/pipeline.pyx b/spacy/pipeline.pyx
index 477c9d6e2..e913b2647 100644
--- a/spacy/pipeline.pyx
+++ b/spacy/pipeline.pyx
@@ -501,15 +501,20 @@ class Tagger(Pipe):
         cdef int idx = 0
         correct = numpy.zeros((scores.shape[0],), dtype='i')
         guesses = scores.argmax(axis=1)
+        known_labels = numpy.ones((scores.shape[0], 1), dtype='f')
         for gold in golds:
             for tag in gold.tags:
                 if tag is None:
                     correct[idx] = guesses[idx]
-                else:
+                elif tag in tag_index:
                     correct[idx] = tag_index[tag]
+                else:
+                    correct[idx] = 0
+                    known_labels[idx] = 0.
                 idx += 1
         correct = self.model.ops.xp.array(correct, dtype='i')
         d_scores = scores - to_categorical(correct, nb_classes=scores.shape[1])
+        d_scores *= known_labels
         loss = (d_scores**2).sum()
         d_scores = self.model.ops.unflatten(d_scores, [len(d) for d in docs])
         return float(loss), d_scores