diff --git a/pipeline/run.py b/pipeline/run.py
index 9a9d5d5..c1bff36 100644
--- a/pipeline/run.py
+++ b/pipeline/run.py
@@ -5,6 +5,44 @@
 import spacy
 from transform_features import get_feature_transformer
 
+def pipeline(args):
+    '''
+    Runs the model loop.
+    '''
+    df = pd.read_csv(args.filename)
+    if args.dedupe:
+        df = df.drop_duplicates(subset='content')
+    if args.reduce:
+        df = restrict_sources(df, column)
+    X = df[args.x_label]
+    y = df[args.y_label]
+    parser = spacy.load('en')
+    X_train, X_test, y_train, y_test = train_test_split(X, y)
+    loop = ModelLoop(X_train, X_test, y_train, y_test, args.models,
+                     args.iterations, args.output_dir,
+                     thresholds = args.thresholds, ks = args.ks)
+    loop.run()
+
+def restrict_sources(df, column, max_size=500, random_state=1):
+    '''
+    Resamples data set such that samples with more than n=500 are re-sampled
+    randomly.
+    '''
+    print("Resampling sources with frequency larger than {}".format(max_size))
+    counts = df.groupby(column).count()
+    counts['count'] = counts[counts.columns[0]]
+    counts = counts.loc[:,'count']
+    to_sample = defaultdict(lambda:set([]))
+    for row in df.itertuples():
+        if counts.loc[row.source] > max_size:
+            to_sample[row.source].add(row[0])
+    remove = []
+    for source in to_sample:
+        size = counts.loc[source] - max_size
+        remove += list(np.random.choice(list(to_sample[source]), size=size, replace=False))
+    df.drop(remove, inplace=True)
+    return df
+
 if __name__=='__main__':
     parser = argparse.ArgumentParser(description='Run a model loop')
     parser.add_argument('filename', type=str,
@@ -25,21 +63,9 @@
                     action="store_true")
     parser.add_argument('--ks', nargs='+', type=float, help='Metrics at k',
                     default = [])
+    parser.add_argument('--reduce', nargs=1, type=int, help='Restrict sample size from large sources',
+                    default = False)
 
     args = parser.parse_args()
     print(args)
-
-    df = pd.read_csv(args.filename)
-    if args.dedupe:
-        df = df.drop_duplicates(subset='content')
-    # print(df.head())
-    X = df[args.x_label]
-    # print(X.head())
-    y = df[args.y_label]
-    # print(y.head())
-    parser = spacy.load('en')
-    X_train, X_test, y_train, y_test = train_test_split(X, y)
-    loop = ModelLoop(X_train, X_test, y_train, y_test, args.models,
-                     args.iterations, args.output_dir, 
-                     thresholds = args.thresholds, ks = args.ks)
-    loop.run()
+    pipeline(args)