diff --git a/downloads-generation/models_class1/generate_hyperparameters.py b/downloads-generation/models_class1/generate_hyperparameters.py
index 65491eafdbd38902096af3d5e3308b40b118b912..a805a32c1f38f938bfb5fc3f08e7d35d689905cb 100644
--- a/downloads-generation/models_class1/generate_hyperparameters.py
+++ b/downloads-generation/models_class1/generate_hyperparameters.py
@@ -64,7 +64,7 @@ base_hyperparameters = {
 
 grid = []
 for train_subset in ["all", "quantitative"]:
-    for minibatch_size in [32]:
+    for minibatch_size in [128]:
         for dense_layer_size in [8, 16, 32, 64]:
             for l1 in [0.0, 0.001]:
                 for num_lc in [0, 1, 2]:
diff --git a/mhcflurry/train_allele_specific_models_command.py b/mhcflurry/train_allele_specific_models_command.py
index eac3f9c3a9fabfeb51a42e3a584b87305e17a383..694e4a47c4e7aed1d4363b3990525371e9ea3f0e 100644
--- a/mhcflurry/train_allele_specific_models_command.py
+++ b/mhcflurry/train_allele_specific_models_command.py
@@ -10,6 +10,7 @@ import traceback
 import random
 from functools import partial
 
+import numpy
 import pandas
 import yaml
 from mhcnames import normalize_allele_name
@@ -472,6 +473,9 @@ def calibrate_percentile_ranks(allele, predictor, peptides=None):
 
 
 def worker_init(keras_backend=None, gpu_device_nums=None):
+    # Each worker needs distinct random numbers
+    numpy.random.seed()
+    random.seed()
     if keras_backend or gpu_device_nums:
         print("WORKER pid=%d assigned GPU devices: %s" % (
             os.getpid(), gpu_device_nums))