Machine Learning and Language Models

This repository contains the code and resources for the machine learning focusing on language models. The project explores the capabilities of GPT-2 and Llama models through pre-training, fine-tuning, and prompting techniques. (please refer to the report.pdf for more details on the project).

Table of Contents

• Part I: Pre-training GPT-2
• Part II: Instruction Fine-tuning GPT-2
• Part III: Chain-of-Thoughts Prompting
• References

Part I: Pre-training GPT-2

In this section, we pre-train the GPT-2 model on a Shakespearean text corpus to generate text in a similar style. The codes are missing for this part, but you can refer to the report.pdf for details on the training and validatin loss, as well as the generated text samples.

• The training and validation loss curves show the model's learning progress (for 5000 steps of iterations).

• The generated text samples demonstrate the model's ability to mimic Shakespearean language, although some grammatical inconsistencies are present due to character-level training.

Part II: Instruction Fine-tuning GPT-2

We fine-tune the pre-trained GPT-2 model using the Alpaca-GPT4 dataset to enhance its instruction-following capabilities.

Alpaca-GPT4 Dataset: The dataset contains 52k instruction-following examples with 1.5M tokens, designed to evaluate the model's ability to follow human instructions.

Instruction Tuning Pipeline: The fine-tuning pipeline uses a specific template for tokenizing instruction-following data

### Instruction: <instruction text here>
### Input: <input text here>
### Response: <response text here>

Memory Efficient Optimization: We experimented with different optimization techniques to reduce memory consumption and computational cost

• AdamW: a variant of Adam optimizer with weight decay regularization
• Stochastic Gradient Descent (SGD): update model using random mini-b (low memory but slower to converge)
• Low-rank Adaptation (LoRA): freezes most model weights and updates only small matrices of parameters (memory-efficient)
• Block Adam (BAdam): uses block-diagonal approximation of the Fisher information matrix (memory-efficient and speed-up convergence)

The fine-tuned models show significant improvements in instruction-following capabilities, with higher relevance and fluency in generated responses.

Part III: Chain-of-Thoughts Prompting

We evaluate the effectiveness of CoT prompts on mathematical benchmarks using the Llama model, using the following datasets for evaluation:

• GSM8K: 8.5k grade-school math word problems (tests basic arithmetic operations)
• NumGLUE: math tasks like arithmetic and logic (tests algebraic reasoning)
• StimuIEq: math equation problems (tests equation-solving)
• SVAMP: complex math word problems (tests problem-solving)

CoT Prompting Strategy and Results:

The CoT prompts are designed to improve the model's reasoning capabilities by providing detailed step-by-step instructions. We evaluate the model's performance with different numbers of CoT examples (0-shot, 2-shot, and 4-shot).

The use of CoT prompts generally improves the model's performance, with 4-shot prompts yielding the best results for most datasets.

References

1. Tom Brown et al. "Language models are few-shot learners". In: Advances in Neural Information Processing Systems. Vol. 33. 2020, pp. 1877-1901.
2. Karl Cobbe et al. "Training verifiers to solve math word problems". In: arXiv preprint arXiv:2110.14168 (2021).
3. Abhimanyu Dubey et al. "The LLaMA 3 herd of models". In: arXiv preprint arXiv:2407.21783 (2024).
4. Diederik P Kingma and Max Welling. "A method for stochastic optimization". In: arXiv preprint arXiv:1412.6980 (2014).
5. Rik Koncel-Kedziorski et al. "Mavps: A math word problem repository". In: Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016, pp. 1152-1157.
6. Swaroop Mishra et al. "Numglue: A suite of fundamental yet challenging mathematical reasoning tasks". In: arXiv preprint arXiv:2204.05660 (2022).
7. Arkil Patel, Satwik Bhattamishra, and Navin Goyal. "Are nlp models really able to solve simple math word problems?" In: arXiv preprint arXiv:2103.07191 (2021).
8. Alec Radford et al. "Language models are unsupervised multitask learners". In: OpenAI blog 1.8 (2019), p. 9.
9. Hugo Touvron et al. "LLaMA 2: Open foundation and fine-tuned chat models". In: arXiv preprint arXiv:2307.09288 (2023).
10. A Vaswani. "Attention is all you need". In: Advances in Neural Information Processing Systems. 2017.