diff --git a/perfkitbenchmarker/linux_benchmarks/mlperf_benchmark.py b/perfkitbenchmarker/linux_benchmarks/mlperf_benchmark.py
index 92e69823ed..c2db362895 100644
--- a/perfkitbenchmarker/linux_benchmarks/mlperf_benchmark.py
+++ b/perfkitbenchmarker/linux_benchmarks/mlperf_benchmark.py
@@ -318,6 +318,7 @@ def _CreateMetadataDict(benchmark_spec):
       'use_tpu': bool(benchmark_spec.tpus),
       'model_dir': benchmark_spec.model_dir,
       'model': benchmark_spec.benchmark,
+      'version': 'v0.6.0',
   }
   if benchmark_spec.tpus:
     metadata.update({
@@ -349,7 +350,6 @@ def MakeSamplesFromOutput(metadata, output, use_tpu=False, model='resnet'):
       r':::MLL (\d+\.\d+) eval_accuracy: {(.*)}', output)
 
   start = None
-  version = 'v0.6.0'
   for wall_time, result in results:
     wall_time = float(wall_time)
     if not start:
@@ -366,7 +366,6 @@ def MakeSamplesFromOutput(metadata, output, use_tpu=False, model='resnet'):
       value = regex_util.ExtractExactlyOneMatch(r'"value": (\d+\.\d+)', result)
     metadata_copy['times'] = wall_time - start
     metadata_copy['epoch'] = int(epoch)
-    metadata_copy['version'] = version
     samples.append(
         sample.Sample('Eval Accuracy',
                       float(value) * 100, '%', metadata_copy))
diff --git a/tests/linux_benchmarks/mlperf_benchmark_test.py b/tests/linux_benchmarks/mlperf_benchmark_test.py
index 687ee24d8c..cdf248072f 100644
--- a/tests/linux_benchmarks/mlperf_benchmark_test.py
+++ b/tests/linux_benchmarks/mlperf_benchmark_test.py
@@ -32,7 +32,7 @@ def setUp(self):
 
   @mock.patch('time.time', mock.MagicMock(return_value=1550279509.59))
   def testTrainResults(self):
-    samples = mlperf_benchmark.MakeSamplesFromOutput({},
+    samples = mlperf_benchmark.MakeSamplesFromOutput({'version': 'v0.6.0'},
                                                      self.contents,
                                                      use_tpu=True,
                                                      model='resnet')