COMP7607: Natural Language Processing - The University of Hong Kong 1 Fall 2024Python

Java Python Assignment 1

COMP7607: Natural Language Processing - The University of Hong Kong

Fall 2024

Due: October 31, 6:00 AM

Recent studies reveal that large language models (LLMs), such as Llama (Dubey et al., 2024) and GPT-4 (OpenAI, 2023), excel in step-by-step reasoning for addressing complex tasks driven by natu- ral language instructions.  In this assignment, you will explore the capabilities of the Llama-3.1-8B- Instruct model in two reasoning domains: mathematics and coding. You can choose one task—either mathematics or coding—based on your interests.

Submit:  You should submit your assignment to the COMP7607 Moodle page.  You will need to submit a PDF file UniversityNumber.pdf of your report (with detailed experimental details) and a zip file UniversityNumber.zip, which includes:

•  .py files.

•  zeroshot.baseline.jsonl

•  fewshot.baseline.jsonl (for math task)

•  [method_name_1] .jsonl

•  [method_name_2] .jsonl

•  [method_combine] .jsonl or  [method_improve] .jsonl

Please note that the UniversityNumber is the number printed on your student card.

1   Introduction

Prompt Engineering refers to methods for how to instruct LLMs for desired outcomes without up- dating model weights.  The core task of assignment 1 is to design methods for prompting LLMs to improve the accuracy of LLMs on math problem-solving or code generation.

Generally, we have two prompting paradigms:

1. Zero-shot prompting is to simply feed the task text to the model and ask for results. For exam- ple, we can send the following message to Llama-3.1-8B-Instruct:

messages  =   [

{"role":    "system",  "content":    [System  Message]},

{"role":    "user",  "content":    "Elsa  has  5  apples.   Anna  has  2 more  apples  than  Elsa . How many  apples  do  they  have  together?"},

]

“{"role":    "system",  "content":    [System  Message]}” can be optional.  [System Message]  is a feature-specific contextual description given to a generative AI model, such as “Your  task  is  to  solve  a  series  of math  word  problems  by  providing  the  final  answer.   Use  the format  ####  [value]  to  highlight  your  answer.”

2. Few-shot prompting presents a set of high-quality demonstrations, each consisting of both in- put and desired output, on the target task. As the model first sees good examples, it can better understand human intention and criteria for what kinds of answers are wanted.  For example, we can send the following 1-shot prompting message to Llama-3.1-8B-Instruct:

messages  =   [

{"role":    "system",  "content":    [System  Message]},

{"role":    "user",  "content":    "Elsa  has  5  apples.   Anna  has  2 more  apples  than  Elsa . How many  apples  do  they  have  together?"},

{  "role":    "assistant",  "content":    "Anna  has  2 more  apples  than  Elsa .   So  Anna  has 2  +  5  =  7  apples.    Elsa  and  Anna  have  5  +  7  =  12  apples  together.    ####  12"  }

{"role":   "user",  "content":    "If  there  are  3  cars  in  the  parking  lot  and  2 more  cars

arrive,  how many  cars  are  in  the  parking  lot?"}, ]

The choice of prompting format, interaction method, and decoding parameters will all significantly influence task performance. Now enjoy exploring the magic of prompting with Llama-3.1-8B-Instruct!

2   Task

For your selected task (math or code), begin by implementing the two simplest baselines: zero-shot and few-shot prompting, using only the problem statement with demonstrations (optional). Please re- fer to the example format in Section 1. Then, choose two advanced methods to enhance performance beyond baselines. We present advanced methods that you may choose to implement in Section 2.1and Section 2.2, respectively. You are also encouraged to propose your own designs!

Here, we provide two surveys that offer a comprehensive overview of the development of prompt- ing engineering and the self-correction of LLMs:

• A survey for prompt engineering for autoregressive language models [Survey Link].

• A collection of research papers for LLMs self-correcting [Survey Link]. Note that we only focus on the “Generation-Time Correction” and “Post-hoc Correction” .

2.1   Mathematical Reasoning

Data   We use the grade school math dataset, GSM8K, which is one of the most popular datasets for evaluating the mathematical reasoning performance of LLMs. Note that the performance of the LLM is measured using “GSM8K/test.jsonl” .  You can utilize “GSM8K/train.jsonl” for other objectives, such as retrieving the most similar examples for the questions in the test file.

Metric   For each method, report the overall accuracy on the 1,319 questions in the test file.  Use the answer extraction function provided in “GSM8K/evaluation.py” for calculating accuracy. Additionally, report the inference cost using two metrics:  wall-clock time and the average number of generated tokens per math question.

Task 1: Implement Baseline    We consider two baselines:

• Zero-shot prompt the model using only the problem statement.

• Few-shot prompts the model with a few demonstrations, each containing manually written (or model-generated) high-quality reasoning chains.

We provide the prompting template in “GSM8K/baseline .py” .

Task 2: Implement two Advanced Prompting Methods   You may explore various advanced strate- gies for improvement, such as retrieving similar demonstrations, decomposing responses, employing response ensemble, engaging in self-improvement, and more.  Below, we present several advanced methods, although not exhaust COMP7607: Natural Language Processing - The University of Hong Kong Assignment 1 Fall 2024Python ive, that you may choose to implement, considering the characteristics of mathematical reasoning tasks. You are also encouraged to propose your own designs!

• Self-Polish: Enhance Reasoning in Large Language Models via Problem Refinement [paper]

• Tree Prompting: Efficient Task Adaptation without Fine-Tuning [paper]

• SELF-REFINE: Iterative Refinement with Self-Feedback [paper]

• PHP: Progressive-Hint Prompting Improves Reasoning in Large Language Models [paper]

• CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing [paper]

• Teaching Algorithmic Reasoning via In-context Learning [paper]

• Contrastive Decoding Improves Reasoning in Large Language Models paper]

• Self-Evaluation Guided Beam Search for Reasoning [paper]

• Skills-in-Context Prompting: Unlocking Compositionalityin Large Language Models [paper]

• Program of Thoughts Prompting:  Disentangling Computation from Reasoning for Numerical Reasoning Tasks [paper]

• Large Language Models are Better Reasoners with Self-Verification [paper]

Task 3: Combine the Two Methods or Enhance One Method   Can you combine the two advanced methods or enhance one method based on your analysis to achieve greater gains? Are complementary or enhanced effects achievable? If not, please explain why.

2.2   Code Generation

A particularly intriguing application of LLMs is code generation,  a task that involves producing source code from natural language descriptions.   This area has garnered substantial interest from both academia and industry, as evidenced by the development of tools like GitHub Copilot.

Data   We use HumanEval leaderboard, which has been established as a de facto standard for eval- uating the coding proficiency of LLMs.  Note that the performance of the LLM is measured using “HumanEval/HumanEval.jsonl” .

Metric   For each method, report the overall accuracy on the 164 questions in the “HumanEval.jsonl” file. We provided “example_problem.jsonl” and “example_solutions.jsonl” under HumanEval to illustrate the format and assist with debugging. You can evaluate the accuracy of the generated code by running:

python  evaluate_functional_correctness .py  –sample_file  example_samples.jsonl  –problem_file example_problem.jsonl

Also, report the inference cost using two metrics:  wall-clock  time and the average number of generated tokens per code problem.

Task 1: Implement Baseline   We use zero-shot prompting as the baseline.  The prompting template is provided in “HumanEval/baseline .py” .

Task 2: Implement two Advanced Prompting Methods   You may explore various advanced strate- gies for improvement.  Below, we present several advanced methods that you may choose to imple- ment. You are also encouraged to propose your own designs!

• Teaching Large Language Models to Self-Debug [paper]

• SelfEvolve: A Code Evolution Framework via Large Language Models [paper]

• Self-Edit: Fault-Aware Code Editor for Code Generation [paper]

• Lever: Learning to verify language-to-code generation with execution [paper]

• Is Self-Repair a Silver Bullet for Code Generation? [paper]

• Codet: Code generation with generated tests [paper]

• SELF-REFINE: Iterative Refinement with Self-Feedback [paper]

• Language agent tree search unifies reasoning acting and planning in language models [paper]

Task 3: Combine the Two Methods or Enhance One Method   Can you combine the two advanced methods or enhance one method based on your analysis to achieve greater gains? Are complementary or enhanced effects achievable? If not, please explain why.

Task 4: Report the Inference Cost

2.3   Model

In this assignment, we will use Llama-3.1-8B-Instruct, which is a powerful open-source model that

natively supports multilingual capabilities, coding, reasoning, and tool usage. For more details about this model, you can refer to the Meta. Blog: https://ai.meta.com/blog/meta-llama-3-1/and https: //www.llama.com/docs/model-cards-and-prompt-formats/llama3_1/ .

3   API

In this assignment, you will interact with the Llama-3.1-8B-Instruct sponsored by SambaNova Sys- tem.  To access this resource, please refer to the instructions in the “SambaNova  Cloud  QuickStart  Guide.pdf” to register and generate your API key. To verify that your API key is functioning correctly, you can either use the provided curl command in the document or run the “test_full_response .py” script.

4   Report

You will write a report including the following parts:

• The description of your implemented methods, including the architecture, the hyperparameters, etc.

• The discussion and table of your results, including the hyperparameters you have tested, the performance of the methods on the test set, the analysis of the results, the advantages compared to the baselines, etc.

5   Note

There are some key points you should pay attention to:

• Your assignment will not be evaluated based solely on the accuracy of the task.  Instead, we primarily assess your work through your experiments and analysis.

• You can adjust the decoding hyperparameters to get a better performance on the test set.  The hyperparameters include the temperature, top-p, top-k, etc.

• The Llama-3 .1-8B-Instruct is the only base model to be used for solving tasks in this assign- ment