Short Answer Questions (20 pts)

1. Question 1 (Data Cleaning)
   Explain why replacing sensitive tokens (URLs, emails, phone numbers, etc.) with semantic placeholders (e.g., [url], [email]) is often better than deleting them outright. Reference the data cleaning goals from Task 1.2 and describe at least two downstream analyses that still benefit from knowing a placeholder existed.
   Experiment Setup for Questions 2 & 3

   Run the following cell to call your generate_news_headline function on both a positive and a negative review across a range of temperature values. Use the outputs to answer the next two questions.

   positive_review = reviews_df['text_clean'].iloc[0]
   negative_review = reviews_df['text_clean'].iloc[2]
   temperature_range = [0.1, 0.2, 0.5, 1.0, 1.5, 2.0, 2.5]
   
   print("--- Positive Review Headlines ---")
   print(f"Original: {positive_review}\n")
   for temp in temperature_range:
       headline = generate_news_headline(positive_review, temperature=temp)
       print(f"Temp {temp:.1f}: {headline}")
   
   print("\n" + "-" * 50 + "\n")
   print("--- Negative Review Headlines ---")
   print(f"Original: {negative_review}\n")
   for temp in temperature_range:
       headline = generate_news_headline(negative_review, temperature=temp)
       print(f"Temp {temp:.1f}: {headline}")

2. Question 2 (Sampling Effects)
   Based on the outputs you generated above, describe how the headlines changed as the temperature increased. Discuss word choice, creativity, factual grounding, and what happened at the highest temperatures. Provide at least two concrete examples pulled from your generations.
3. Question 3 (Temperature Selection)
   Given your experiment, propose an "ideal" temperature range for creative-but-accurate headline generation. Cite three specific LLM tasks that demand low temperatures and three that benefit from higher temperatures, explaining why temperature matters for each case.
4. Question 4 (Comparing LLMs)
   Using the evaluation framework from Part 3, compare and contrast at least two Hugging Face checkpoints you profiled. Reference quantitative metrics (accuracy, latency, or other observations from evaluate_model) and qualitative behaviors. Discuss when you would choose one model over the other.