Homework 3: Prompting Engineering and Ethics

Learning Objectives

• Recall how to evaluate generated output
• Identify what prompting techniques produce better output
• Determine when smaller LLMs like Mistral/Llama-2/etc. would be worth using
• Consider the ethical implications of using an LLM

Part 1: Prompting

Helpful Resources

• Original paper on few-shot prompting: Language Models are Few-Shot Learners
• Chain-of-thought prompting: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
• Prompting slides: 26-04-09_prompting.pdf

Other ways of prompting

• Prompt-and-Rerank: A Method for Zero-Shot and Few-Shot Arbitrary Textual Style Transfer with Small Language Models
• Cutting Down on Prompts and Parameters: Simple Few-Shot Learning with Language Models

What to do

Start with this notebook and change the prompts of the model to answer the questions below. This notebook also has the data. Any time we ask for a prompt, please be sure to keep all the cells in the notebook with your prompt text. Copy the output from the model into the document where you answer the questions below. (This will keep the output in case the notebook is accidentally rerun.) The number of suggested prompts are minimums.

1. (2 points for explaining why your evaluation metric is reasonable) You will need some objective evaluation metrics to determine how well your prompting is doing. You will be using two objective evaluation metrics throughout this homework. Take the time now to implement both in your code. a. The for the first one, you will use BLEU. BLEU is common evaluation metric used in NLP created to evaluate how accurate machine translation (the computational translation of one human language to another) systems were. You can call BLEU using this library: https://www.nltk.org/api/nltk.translate.bleu_score.html b. For the second evaluation metric, you will pick your own. If you find a method that wasn’t mentioned in the class slides, please cite your source!
2. First, try the generation task using zero-shot prompting in the plainest way possible. Just ask the model to do the task. Don’t give any examples for how to do it, don’t use any fancy prompting techniques. Just ask it like you’re asking a human to do the task. We will refer to this as your baseline.
   a. (2 pts) Provide 2 prompts. b. (2 pts) Run your 2 implemented objective evaluation metrics and report your scores.
3. In class, we talked about providing a “role” to the model as part of the instructions.
   a. (2 pts) Provide 2 prompts trying this out.
   b. (2 pts) How does this compare to the output from the baseline? Use both objective measures from question 1 and also use your intuition for a more “subjective measure”.
4. What happens when you use few-shot prompting?
   a. (4 pts) Provide 4 prompts trying out different numbers of examples (i.e., different “n-shot” prompts).
   b. (2 pts) How does changing the number of examples affect performance? Use your objective measure and your subjective measure.
   c. (1 pt) Is there a cut-off where it isn’t helpful to have more examples?
5. Consider chain-of-thought prompting, where you get the model to “show its work” to produce better results.
   a. (2 pts) Provide 2 prompts trying chain-of-thought prompting.
   b. (2 pts) How does this compare to your baseline? Use your objective measures and your subjective measure.
6. Try two other techniques that we talked about in class. a. (2 pts) Provide both of those prompts.
   b. (2 pts) How do these compare to your baseline? Use your objective measures and your subjective measure.
7. These models are not considered “state of the art”, but they could still have their uses. Recall some of the tasks you looked into in HW1.
   a. (3 pts) What types of NLP tasks do you think the model you chose would be good at? Why?
   b. (3 pts) What would it be bad at? Why?
   c. (2 pts) Was the model you chose good at this task? Why do you think that?
8. (2 pts) What types of “everyday” tasks (e.g., doing taxes, writing code) would the model you chose be good at? Why?

Extra credit

• Play around with different sampling strategies. Use this guide for implementing them using Hugging Face. Pick one of the prompts that you used above and keep it the same as you change to different ways of sampling. Try 3 different ways of sampling.
  a. (1 pt) What are the different ways you tried?
  b. (3 pts, 1 pt each) How does it affect the generation?

Part 2: Ethics

Although you are using a smaller model than some state-of-the-art models and you’re using a quantized (condensed) version of it, there can still be some ethical ramifications. In this part, you will be considering some of these.

1. (2 points) Using this calculator, approximate how much energy you used for this homework. Select the model you used (or the closest relative to the model you used) and enter in how many tokens the prompts generated across all of the prompts you used whether or not you reported them in the homework. If you called the model, that adds to your total.
   Under “Environmental Impact”, select the “Grid Region” of “United States (Average)”. Report your total Carbon Emissions.
   The average household in the US generates 0.345 kg CO2 from electricity in a year. Compare this number to the number you calculated.
2. (2 points) Were there any toxic generations to any of the prompts you provided for Part 1? Discuss them if so. If not, why do you think you didn’t see any?
3. (2 points) Considering how the model did across your prompts, do you believe that it was trained on WikiHow data? Why or why not?
4. (2 points) Regardless of your answer to the previous question, let’s say the model was trained on WikiHow data. How might you mitigate the impact?

What to turn in

• The code that you used for evaluation
• A document with the answers to your questions, all of your prompts, and the output from the prompts

Grading