Intuition behind action predicted by Contextual Bandit on Training dataset

Question

Doing the Contextual Bandit tutorial https://vowpalwabbit.org/docs/vowpal_wabbit/python/latest/tutorials/python_Contextual_bandits_and_Vowpal_Wabbit.html I wanted to try what would the predicted choice/action be when the training dataset was passed to get predictions from the model. I got the following results and I don't understand why the model predicted a different action and not the exact same action as in the training dataset.

I get the predictions like mentioned in the tutorial for the Test DataFrame.

From an expectation point of view, I think the model should have given the same actions as the actions in the training dataset when asked to predict on it.

Code reference

# Prepare the training dataset.
train_data = [
    {
        'action': 1,
        'cost': 2,
        'probability': 0.4,
        'feature1': 'a',
        'feature2': 'c',
        'feature3': '',
    },
    {
        'action': 3,
        'cost': 0,
        'probability': 0.2,
        'feature1': 'b',
        'feature2': 'd',
        'feature3': '',
    },
    {
        'action': 4,
        'cost': 1,
        'probability': 0.5,
        'feature1': 'a',
        'feature2': 'b',
        'feature3': '',
    },
    {
        'action': 2,
        'cost': 1,
        'probability': 0.3,
        'feature1': 'a',
        'feature2': 'b',
        'feature3': 'c',
    },
    {
        'action': 3,
        'cost': 1,
        'probability': 0.7,
        'feature1': 'a',
        'feature2': 'd',
        'feature3': '',
    },
]
train_df = pd.DataFrame(train_data)

train_df['index'] = range(1, len(train_df) + 1)
train_df = train_df.set_index('index')


# Prepare the test dataset.
test_data = [
    {'feature1': 'a', 'feature2': 'c', 'feature3': ''},
    {'feature1': 'b', 'feature2': 'c', 'feature3': ''},
    {'feature1': 'a', 'feature2': '', 'feature3': 'b'},
    {'feature1': 'b', 'feature2': 'b', 'feature3': ''},
    {'feature1': 'a', 'feature2': '', 'feature3': 'b'},
]
test_df = pd.DataFrame(test_data)

test_df['index'] = range(1, len(test_df) + 1)
test_df = test_df.set_index('index')


# Initialize a contextual bandit with four possible actions.
model = vowpalwabbit.Workspace('--cb 4', quiet=True)


# Train the model on each of the training samples.
for i, row in train_df.iterrows():

    # Construct the example in the required Vowpal Wabbit format.
    row_vw = ('{}:{}:{} | {} {} {}'.format(
        str(row['action']),
        str(row['cost']),
        str(row['probability']),
        str(row['feature1']),
        str(row['feature2']),
        str(row['feature3']),
    ))
    model.learn(row_vw)


# Get prediction from the model on both the training and test dataset.
def calculate_prediction(row, model):

    # Construct the example in the required Vowpal Wabbit format.
    row_vw = '| {} {} {}'.format(row['feature1'], row['feature2'], row['feature3'])

    # Get prediction from the model.
    predicted_action = model.predict(row_vw)
    return predicted_action

train_df['predicted_action'] = train_df.apply(calculate_prediction, axis=1, model=model)
test_df['predicted_action'] = test_df.apply(calculate_prediction, axis=1, model=model)


# Predicted actions on the train dataset.
train_df