HKU DASC-7606 Assignment 2 NLP: In-context Learning with Large Language Models
For questions and discussion,
- We encourage you to use GitHub Issues of this repository.
- Or if you prefer online doc: Discussion doc.
This codebase is only for HKU DASC 7606 (2023-2024) course. Please don't upload your answers or this codebase to any public platforms (e.g., GitHub) before permitted. All rights reserved.
Large language models, such as GPT-3 and its successors, represent a significant leap forward in the field of artificial intelligence, specifically within the realm of natural language processing. These models are trained on vast datasets encompassing a wide swath of human knowledge, allowing them to generate text that can mimic human writing styles with remarkable accuracy. A key feature of these models is their ability to perform "in-context learning", which means they can understand and respond to new instructions or information presented within the immediate context of their input text. This capability enables a form of dynamic adaptation to new tasks without the need for explicit retraining or fine-tuning on specific datasets. As a result, large language models have found applications across a diverse range of fields, from automated content creation to sophisticated dialogue systems, showcasing their versatility and the power of advanced neural network architectures.
The task of multiple-choice question answering is a vital tool in assessing the reasoning capabilities of large language models. It involves presenting the model with a question alongside several possible answers, only one of which is correct. This approach is particularly beneficial as it gauges a model's ability to comprehend, deduce, and apply knowledge acquired during its training to new situations. Notably, initiatives like the Massive Multitask Language Understanding (MMLU) and the AI2 Reasoning Challenge (ARC) are instrumental in this context. They are designed to challenge a wide array of knowledge and reasoning skills, spanning science and mathematics to literature and common sense reasoning. For this assignment, we exclusively utilize the ARC dataset, which allows for rigorous, automatic evaluation of a model's reasoning capabilities.
In this assignment, you will gain valuable knowledge and practical experience in several key areas:
-
Understand the architecture of large language models by focusing on the deployment of a 1.5 billion parameter model in a local environment.
-
Delve into the concept of in-context learning, learning how to harness this powerful technique to adapt the model to new tasks. You will explore how to make in-context learning work effectively by providing the model with examples or instructions within the input text, guiding it to understand and perform tasks without explicit retraining.
-
Understand the format of multiple-choice question answering tasks. This includes how questions are presented along with several answer choices, challenging the model to use its reasoning abilities to select the correct answer.
-
Learn how to apply in-context learning techniques to multiple-choice question answering problems.
Participants may engage with the assignment through two primary environments: locally on their personal machine or remotely on the HKU GPU Farm.
Note: After completing these setup steps, proceed directly to creating the Python environment, omitting the local setup instructions.
For the duration of this course, you are granted access to the HKU GPU Farm for assignment purposes. It is advised to consult the quickstart guide available on the official website to acquaint yourself with the HKU GPU Farm's environment.
Ensure you have acquired the following competencies after reviewing the quickstart documentation:
- Proficiency in accessing the GPU Farm and utilizing GPUs in interactive mode, which is highly recommended for this assignment to expedite your experiments.
- Familiarity with initiating Jupyter Lab sessions without launching a web browser.
- Knowledge of managing tmux sessions to mitigate issues arising from unstable network connections.
If you have GPU resources on your personal computer, preparing for local work requires installing GPU drivers, CUDA, cuDNN, and PyTorch. Although completing the assignment without a GPU is feasible, model inference will be significantly slower. Therefore, having a GPU offers a substantial advantage for enhancing performance throughout the assignment.
Installing Python: The code for this assignment has been verified with Python version 3.10.9.
Virtual Environment: The use of a virtual environment via Anaconda is recommended for this project. If you decide against using a virtual environment, it is your responsibility to ensure all dependencies are installed. To establish a Conda virtual environment, execute the following commands:
conda create -n nlp_env python=3.10.9
conda activate nlp_env
Follow the official PyTorch installation guidelines to set up the PyTorch environment. This guide uses PyTorch version 2.0.1 with CUDA 11.7. Alternate versions may be specified by adjusting the version number:
pip install torch==2.0.1
Proceed to install sentence_transformers, transformers, accelerate, and str2bool:
pip install sentence_transformers==2.3.1
pip install transformers==4.31.0
pip install accelerate==0.22.0
pip install str2bool
In-context learning enables large language models to adapt to and perform new tasks based on the examples and instructions embedded directly within the input prompt. This feature exploits the model's extensive pre-training on diverse text to infer the task requirements from the given context and generate a suitable output without additional training.
To illustrate, let's consider teaching a model to answer simple arithmetic questions through in-context learning. Below is a complete prompt that includes contextual cues (examples of arithmetic operations and their answers) and a target question for the model to answer:
Input Prompt:
- Example 1: What is 3 plus 2? Answer: 3 plus 2 is 5.
- Example 2: What is 6 minus 1? Answer: 6 minus 1 is 5.
- Question: What is 5 plus 7?
In this prompt, the first two lines are the contextual cues, demonstrating how the model should interpret and respond to arithmetic questions. The "Example 1" and "Example 2" parts show the model the format of the question and the expected format of the answer. The line starting with "Question" is the target question we want the model to answer, based on the pattern it learned from the contextual cues. By analyzing the input prompt, the model learns to map a new arithmetic input ("What is 5 plus 7?") to the appropriate output, leveraging the format and knowledge presented in the examples.
To prevent data leakage and maintain the validity of evaluations in in-context learning, it's important that the examples provided as contextual cues are not taken from the test set. These examples should instead be selected from the training dataset or be manually created. Selecting examples carefully is essential for effectively applying the model's extensive pre-trained knowledge.
The multiple-choice question answering task involves presenting a model with a question and a set of possible answers, among which only one is correct.
To illustrate, consider the following example:
Question: George wants to warm his hands quickly by rubbing them. Which skin surface will produce the most heat?
Candidate Answers: (A) dry palms (B) wet palms (C) palms covered with oil (D) palms covered with lotion
Gold Answer: A
In this example, the model is tasked with analyzing the question and the provided candidate answers to determine which option correctly answers the question.
Throughout this assignment, you'll employ in-context learning to boost question-answering accuracy using the ARC dataset. Key to success is the adept formatting of examples and the strategic selection from the training data, ensuring the model effectively leverages its pre-trained knowledge.
The provided codebase is structured to facilitate conducting experiments with the ARC dataset. Here's an overview of the directory and file organization:
└── src
├── data
│ ├── ARC-Challenge-test.jsonl # Test set for the ARC Challenge
│ ├── ARC-Challenge-train.jsonl # Training set for the ARC Challenge
│ ├── ARC-Challenge-validation.jsonl # Validation set for the ARC Challenge
│ ├── ARC-Easy-test.jsonl # Test set for the ARC Easy
│ ├── ARC-Easy-train.jsonl # Training set for the ARC Easy
│ └── ARC-Easy-validation.jsonl # Validation set for the ARC Easy
├── cache_utils.py # Utility functions for caching
├── modeling_attn_mask_utils.py # Utilities for attention mask modeling
├── download.py # Script to download necessary models
├── configuration_phi.py # Defines the large language model (LLM) configuration
├── modeling_phi.py # Implements the LLM architecture
├── tokenization_codegen.py # Tokenizer definition for the model
├── eval_fewshot.py # Main script for conducting experiments
├── eval_fewshot_multigpu.py # Supports experiments using multiple GPUs
├── acc.py # Calculates accuracy metric
└── README.md
To run experiments and compute the accuracy metric, the following commands can be used:
python eval_fewshot.py --data_path "data/ARC-Easy-validation.jsonl" --device_id "0,1" --model path_of_llm --embedder path_of_embedder --start_index 0 --end_index 9999 --max_len 1024 --output_path path_to_save_model_predictions --overwrite False --prompt_type "v2.0" --N 8 --top_k True --top_k_reverse False
python acc.py --prediction_path path_to_save_model_predictions
Parameters explained:
modelandembedder: Specify paths to the large language model (e.g., Phi-1.5) and the embedding model (e.g., bge-small-en-v1.5), respectively. These models can be prepared with thedownload.pyscript.data_path: Path to the JSONL file containing the dataset for evaluation.device_id: GPU device IDs for running the experiments.max_len: Maximum sequence length for the model input.output_path: Directory to save model predictions.prompt_type: Version of the prompt formatting.N: Number of examples to include in the prompt.top_k: Whether to select the top k similar examples based on embedding similarity for in-context learning.top_k_reverse: Determines the order of appending similar examples, where the most similar example is placed closest to the question at hand.
Use eval_fewshot_multigpu.py in place of eval_fewshot.py for splitting the workload across multiple GPUs if running the model on a single GPU is not feasible.
Task 1: Implement the Self-Attention Function
Task 2: Implement the Residual Connection
Task 3: Implement the Feed-Forward Layer
The provided codebase for the Phi-1.5 model is nearly complete, except for specific sections marked with Write Your Code Here in modeling_phi.py, where you are to implement the missing functionalities.
Task 4: Implement the Example Formatting Function
This function converts raw examples into a format suitable for in-context learning. Completion is required in sections tagged with Write Your Code Here within eval_fewshot.py and eval_fewshot_multigpu.py.
Task 5: Implement Code to Select the Top K Similar Examples
Finish implementing the generate_prompt function in both eval_fewshot.py and eval_fewshot_multigpu.py to select examples based on embedding similarity.
Task 6: Implement Code to Calculate the Log-Likelihood of an Answer Choice
Complete relevant sections within eval_fewshot.py and eval_fewshot_multigpu.py to calculate the likelihood of each answer choice.
Task 7: Predict Outputs for the Test Set
Predict the outputs for the test set, which will be released seven days before the deadline. You are encouraged to explore advanced models and hyper-parameter tuning to enhance model performance.
Task 8: Write a Report
Your report should comprise an introduction, methodology, and experiments section. Detail how you improved the model, considering options like:
- Hyper-parameter Tuning: Explore critical hyper-parameters like the number of examples in the prompt, and the selection of top k similar examples based on embedding similarity.
- Pre-trained Models for Embedding: Experiment with advanced neural networks or customize your own for embedding demonstrations and questions.
- Formats of In-Context Demonstrations: Experiment with different demonstration formats and augment questions with relevant external knowledge, inspired by this paper.
- Expanding Context Length: Address limitations on context length by including more demonstrations or modifying embeddings for longer sequences, drawing on this paper.
In the method section, describe the approaches you used to improve the baseline model performance. In the experiment section, analyze the dataset statistics, qualitative evaluations of the model's performance, and case analysis.
-
Final Report (PDF, up to 4 pages)
1.1 Introduction. Task overview, background, and related works.
1.2 Methods. Improvements to the baseline model, including but not limited to the methods above.
1.3 Experiments & Analysis (IMPORTANT) Analysis is the most important part of the report. Possible analysis may include but is not limited to (i) Dataset analysis (dataset statistics) (ii) Qualitative evaluations of your model. Select several specific cases in the dataset and see if your model correctly finds the answer. Failure case analysis is also suggested. (iii) Ablation studies. Analyze why better performance can be achieved when you made some modifications. The performance on the validation set should be given to validate your claim.
-
Codes
2.1 All the codes.
2.2 README.txt that describes which python file has been added/modified.
-
Model Predictions on the Test Set (Submit in a zip file)
If your student id is 30300xxxxx, then the file should be organized as follows:
30300xxxxx.zip
|-report.pdf
|-src
| |-README.md
| |-your code
|-model_predictions.zip
The deadline is Apr.5 (Fri).
Late submission policy:
- 10% for late assignments submitted within 1 day late.
- 20% for late assignments submitted within 2 days late.
- 50% for late assignments submitted within 7 days late.
- 100% for late assignments submitted after 7 days late.
For any questions about the assignment which potentially are common to all students, you shall first look for related resources as follows,
- We encourage you to use
GitHub Issuesof this repository. - Or if you prefer online doc: Discussion doc.
For any other private questions, please contact Xueliang Zhao (xlzhao22@connect.hku.hk) via email.
The assessment of submissions will be conducted with a focus on both the performance metrics achieved on the ARC-Easy-test and ARC-Challenge-test datasets and the comprehensiveness of the submitted report. TAs will evaluate the accuracy of the predicted answers as part of the review process.
The evaluation criteria are divided into two primary components: (1) model performance on the ARC-Easy-test and ARC-Challenge-test datasets, and (2) the quality of the final report, with the latter accounting for 30% of the total marks:
-
Model Performance (70% of total marks): Marks will be given based on the weighted performance of your model across both test sets, calculated as (0.3 \times \text{acc_easy} + 0.7 \times \text{acc_challenge}), where (\text{acc_easy}) represents the accuracy on the ARC-Easy test set and (\text{acc_challenge}) denotes the accuracy on the ARC-Challenge test set:
- A combined score above 63% will get the full mark of this part.
- Scores between 60-63% will get 90% mark of this part.
- Scores between 57-60% will get 80% mark of this part.
- Scores between 54-57% will get 70% mark of this part.
- Scores between 51-54% will get 60% mark of this part.
- Scores above 51% will get 50% mark of this part.
- Others will get 0% mark.
-
Final Report (30% of total marks): The marks will be given mainly based on the richness of the experiments & analysis.
- Rich experiments + detailed analysis: 90%-100% mark of this part.
- Reasonable number of experiments + analysis: 70%-80% mark of this part.
- Basic analysis: 50%-60% mark of this part.
- Not sufficient analysis: lower than 50%.
- Textbooks Are All You Need II: phi-1.5 technical report
- Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge
- UNIFIEDQA: Crossing Format Boundaries with a Single QA System
- Extending Context Window of Large Language Models via Positional Interpolation