- Tinghao (Gavin) Zhang - tz2769
- Haonan (Gary) Guo - hg2924
This project aims to address the challenge of verifying answers to math problems. The core task is to fine-tune a large language model (Llama-3.1-8B) to accurately determine whether a given solution to a math problem is correct. The model's final output should be a simple boolean value: True (correct) or False (incorrect).
This project is part of the "Math Question Answer Verification Competition," and all experiments and code are centered around the goals of this competition.
This project utilizes a public dataset from Hugging Face:
- Dataset Name:
ad6398/nyu-dl-teach-maths-comp - Content: The dataset contains a large number of math questions, corresponding solutions, and a label (
is_correct) indicating whether the solution is correct.
In the various notebooks, we have partitioned the full dataset into different sizes for training and validation sets based on our experimental needs.
To efficiently fine-tune a large model on limited computing resources (like the free tier of Google Colab), we employed the following key technologies:
- Model:
unsloth/Meta-Llama-3.1-8B, the only model specified for this competition. - Efficient Fine-tuning Library:
unsloth, which makes it possible to fine-tune large models on a single GPU by optimizing memory usage and computation speed. - 4-bit Quantization: Compressing the model's weights from standard 32-bit or 16-bit floating-point numbers to 4-bit integers, which significantly reduces GPU memory consumption.
- LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning technique. It speeds up training and reduces resource usage by adding small, trainable "adapter" matrices alongside the model's existing layers, rather than training the entire model's billions of parameters.
This project includes several Jupyter Notebook files that document the different stages and experimental iterations of our problem-solving process.
- Description: This is the initial starter script for the project, designed to quickly run through the entire workflow.
- Configuration: Uses a smaller subset of the data (5,000 training samples and 500 validation samples) and basic training parameters.
- Objective: To serve as a baseline model and verify the correctness of all stages, including data loading, model training, and submission file generation.
- Description: An extended experiment based on the first version.
- Configuration: Increased the amount of training data (10,000 training samples and 1,000 validation samples) and made initial adjustments to hyperparameters such as learning rate and batch size.
- Objective: To explore the impact of larger-scale data on model performance.
- Description: These two notebooks are the advanced versions at the core of the project, introducing more robust training and evaluation mechanisms.
- Configuration:
- Used 10,000 training samples and 1,000 validation samples.
- Adopted an epoch-based training approach to allow the model to learn more thoroughly.
- Introduced validation set evaluation during training (
eval_strategy = "steps"), enabling real-time monitoring of the model's performance on unseen data and saving the best model. - Implemented more detailed training logs and performance reports (achieving validation accuracies of 82.1% and 81.8%, respectively).
- Objective: To find a better-performing model through systematic evaluation and hyperparameter tuning.
- Description: A highly integrated and compact final version script.
- Configuration:
- Used a larger dataset (50,000 training samples and 5,000 validation samples).
- Employed a different prompt format.
- Consolidated all steps (from environment setup to submission file generation) into a single code cell for easy execution and reproducibility.
- Objective: To train on a larger scale and serve as a final, one-click runnable solution.
The link to the trained model weights can be found in the ModelWeight_Link.md file.
The final submission file submitted to Kaggle is submission 0.83.csv, which achieved a score of 0.83 on the competition leaderboard.