🚀 Master's Thesis Experimentation Framework

👋 Overview

This repository contains the code used in my Master's Thesis on data poisoning attacks for models like Phi3, CodeT5, and CodeGen.

Feel free to use, modify, or expand upon it — just please tag or cite my work once my thesis is published! 😊

I’d love to know if this code helped you, whether it was to expand my research, prove something wrong, or for any other reason. 🎯

---

🗂 Folder Structure

Here’s an overview of the folder structure used for my experiments:

• datasets/: Contains all datasets (raw, filtered, and poisoned), divided into training, test, and validation sets.

• DNT/ (Do Not Touch): Stores logs and results from the experiments carried out during the thesis. It contains critical data that should not be modified, such as detailed experiment output logs.

• experiment_results/: Includes files for output samples, success rates, and the parameters used in each experiment. This folder serves as a comprehensive record of the experimental conditions and outcomes.

• results/: Contains files related to GPU VRAM usage, power consumption, and the computational cost of running the experiments. Useful for understanding resource usage during the naturalization and poisoning processes.

---

🛠️ Execution Order

To replicate my experiments, run the code in the following order:

1. 01_DATASET_CLEANING.ipynb
   Cleans and downloads the dataset.

2. 02_DATASET_NATURALIZATION_nonNLP.ipynb
   Performs naturalization (poisons the dataset and sets industry percentages).

3. MONITOREO.ipynb
   Tracks GPU VRAM usage and power consumption and gives you an estimated cost based on the electricity price of your choosing.

After that, feel free to run any of the experiments provided for Phi3, CodeT5, and CodeGen.

Note: These experiments are quite barebones and can definitely be improved! 🔧

---

---

⚡ Naturalization Details

• 02_DATASET_NATURALIZATION.ipynb uses LLaMA3 for naturalization:
  • It takes around 22 days on an RTX 4090 to naturalize 251k rows.
  • The standard method used in 02_DATASET_NATURALIZATION_nonNLP.ipynb takes around 15 seconds to execute on the same hardware.

---

🔗 Thesis Link

Once my thesis is published, I’ll link it here. Stay tuned! 📖

---

🙌 Citation

If you use this code in your research or projects, please cite my work or tag me — I'd love to see how it helped! 🌟