diff --git a/unimore_bda_6/__main__.py b/unimore_bda_6/__main__.py
index 046ac3f..d492b6b 100644
--- a/unimore_bda_6/__main__.py
+++ b/unimore_bda_6/__main__.py
@@ -65,6 +65,10 @@ def main():
                         run_counter += 1
                         slog.debug("Run #%d", run_counter)
 
+                        if run_counter >= 100:
+                            slog.fatal("Exceeded 100 runs, giving up and exiting...")
+                            exit(2)
+
                         try:
                             slog.debug("Instantiating %s with %s...", SentimentAnalyzer.__name__, Tokenizer.__name__)
                             sa = SentimentAnalyzer(tokenizer=Tokenizer())
diff --git a/unimore_bda_6/analysis/base.py b/unimore_bda_6/analysis/base.py
index 64dee81..d4d0c19 100644
--- a/unimore_bda_6/analysis/base.py
+++ b/unimore_bda_6/analysis/base.py
@@ -50,8 +50,11 @@ class BaseSentimentAnalyzer(metaclass=abc.ABCMeta):
         for review in evaluation_dataset_func():
             resulting_category = self.use(review.text)
             evaluated += 1
-            correct += 1 if round(resulting_category) == round(review.category) else 0
-            score += 1 - (abs(resulting_category - review.category) / 4)
+            try:
+                correct += 1 if round(resulting_category) == round(review.category) else 0
+                score += 1 - (abs(resulting_category - review.category) / 4)
+            except ValueError:
+                log.warning("Model execution on %s resulted in a NaN value: %s", review, resulting_category)
 
         return EvaluationResults(correct=correct, evaluated=evaluated, score=score)
 
diff --git a/unimore_bda_6/analysis/tf_text.py b/unimore_bda_6/analysis/tf_text.py
index 1e6066c..351396c 100644
--- a/unimore_bda_6/analysis/tf_text.py
+++ b/unimore_bda_6/analysis/tf_text.py
@@ -248,7 +248,7 @@ class TensorflowPolarSentimentAnalyzer(TensorflowSentimentAnalyzer):
 
         log.debug("Compiling model: %s", model)
         model.compile(
-            optimizer=tensorflow.keras.optimizers.Adam(global_clipnorm=1.0),
+            optimizer=tensorflow.keras.optimizers.Adadelta(global_clipnorm=1.0),
             loss=tensorflow.keras.losses.MeanSquaredError(),
             metrics=[
                 tensorflow.keras.metrics.MeanAbsoluteError(),