Build Naive Bayes spam classifier with F1

Suppose you are working in a Jupyter notebook with a labeled binary text classification dataset for separating messages into two categories: spam and not_spam.

Your implementation should do the following:

• Perform basic text preprocessing; simple cleaning and tokenization are enough.
• Convert the cleaned text into numeric features that a Naive Bayes model can use, such as a bag-of-words representation or TF-IDF.
• Train a Naive Bayes classifier on the resulting features.
• Evaluate the model using the F1 score, and clearly state which variant is reported: the F1 score for the positive class (spam), or macro/micro averaging.
• Apply the trained model to several test examples and display the predicted labels, and optionally the predicted probabilities.

Keep these requirements in mind:

• The class distribution may be imbalanced.
• Avoid data leakage: fit the text vectorizer or feature transformation only on the training split.
• If only one labeled dataset is provided, you may split it into training and validation portions in a reasonable way.

Any correct output showing the requested information is accepted; the exact F1 value and predictions depend on your model and data split.

Example 1:

Input:
train_texts = ["Limited offer, claim your prize now", "Can we move our sync to 3pm?", "Weight loss pills online", "I'll review the document tonight"]
train_labels = ["spam", "not_spam", "spam", "not_spam"]
test_texts = ["You have been selected for a reward", "Please share the agenda before the call"]

Output:
F1 score (positive class: spam): 0.67
Predicted labels: ["spam", "not_spam"]

Explanation: The first test message is classified as spam, while the second is classified as not_spam.

Example 2:

Input:
train_texts = ["URGENT: update your payment details", "Lunch plans still on?", "Free money guaranteed", "The report is attached"]
train_labels = ["spam", "not_spam", "spam", "not_spam"]
test_texts = ["Verify your account immediately", "Thanks for sending the notes"]

Output:
F1 score (positive class: spam): 0.50
Predicted labels: ["spam", "not_spam"]
Predicted probabilities (spam): [0.92, 0.18]

Explanation: The displayed probabilities are optional and show the model's confidence that each test message belongs to the spam class.

Constraints:

• The dataset is binary: every label is either spam or not_spam.
• The class labels may be imbalanced.
• The feature extraction process must be fitted only on the training split.
• A train/validation split may be created if only one labeled set is available.
• Report exactly which F1 score variant is used.