SURGE: Exploring LLMs, Detection, and GAN-Augmented Models

A SURGE through the world of large language models (LLMs), text-detection, and GAN-based attention methods.

Repository: https://github.com/HARSHITJAIS14/SURGE

---

🗂️ Repository Structure

harshitjais14-surge/
├── README.md
├── Week1/
│   └── LLMIntro-2024-07-22-1743.excalidraw
├── Week2/
│   └── Code/
│       ├── mpi.ipynb
│       └── mpi\_120.csv
├── Week3/
│   └── LLMTextDetectionSurvey.pdf
├── Week4/
│   ├── bert-pretrainingdetector.ipynb
│   ├── description.md
│   └── merged\_dataset.csv
├── Week6/
│   ├── gan-detection.ipynb
│   └── gan\_bertattention.ipynb
└── Week7/
└── GANBERT\_pytorch.ipynb

---

📖 Week-by-Week Summary

Week 1: Introduction to LLMs

In Week 1, we dove into the fundamentals and workflow of large language models—from pretraining to supervised fine-tuning, and finally to reinforcement learning (including RLHF). Along the way, we covered key concepts such as base architectures, dataset creation, instruction tuning, and common pitfalls like hallucinations. We also did a hands-on quickstart with the OpenAI API.

Topics Covered:

• LLM architecture & pretraining stages
• Supervised Fine-Tuning (SFT)
• Reinforcement Learning & RLHF
• Prompting basics
• Hallucination in LLMs
• OpenAI API usage

Resources:

1. Busy Person's Intro to LLMs (YouTube)
2. Deep Dive into LLMs (YouTube)
3. Basics of Prompting (YouTube)
4. OpenAI API Quickstart Guide

Artifacts:

• Diagrammed workflow in Excalidraw: Week1/LLMIntro-2024-07-22-1743.excalidraw

Week 2: MPI Personality Evaluation

• Implemented a Machine Personality Inventory tool using Big Five Personality Factors (OCEAN) for assessment of ChatGPT model on text data.
• Notebook mpi.ipynb walks through data preprocessing, feature extraction, and personality inference.
• Dataset sample in mpi_120.csv.

Week 3: Survey on LLM Text Detection

• Read a literature survey of LLM text-detection approaches (PDF in Week3) which gave a full view over the works done over LLM Text Detection till 2023.
• Summarized methods, benchmarks, and open challenges in a concise report and added it to the merged pdf.

Week 4: BERT-Based Pretraining Detector

• Developed a detector to distinguish pretrained vs. fine-tuned text using BERT.
• Notebook bert-pretrainingdetector.ipynb includes model training and evaluation.
• merged_dataset.csv is a smaller version of the CHEAT dataset; description.md details dataset construction.

Week 5: Exploring Datasets

• Learnt about a lot of dataset for LLM Text Detection and listed some of the major datasets and their papers in the Week5 Directory.

Week 6: GAN-Based Detection & Attention

• Explored GAN-based methods for generating and detecting synthetic text/images.
• gan-detection.ipynb builds a basic GAN over word co-occurence matrix as shown in the pdf of the research paper gan_detection_compressed.pdf for adversarial examples.
• gan_bertattention.ipynb adds a BERT-attention module to enhance detection robustness.

Week 7: GAN-BERT in PyTorch

• Full PyTorch implementation of GANBERT paper: integrating GAN-generated data into BERT training loops.
• Notebook GANBERT_pytorch.ipynb demonstrates training and performance analysis over the short version of RAID dataset.

---