Posholi Nyamane

Large Language Models(LLMs) Question Answering System

Developed an intelligent question-answering system using Python, Pytorch, and Large Language Models (LLMs). The system features a comprehensive text preprocessing module, meticulous data analysis, advanced Natural Language Processing (NLP) techniques, and is powered by a finetuned GPT-2 model. Designed for scalability and effectiveness, the system demonstrates significant potential for real-world applications, such as automating customer support services.

Data preprocessing:

• Built a comprehensive preprocessing module to clean and prepare text: Developed a robust preprocessing pipeline for text data that involved techniques such as tokenization, stemming, and removal of stopwords and special characters, setting the stage for effective data analysis.

Data Analysis:

• Understanding Data Distribution: Conducted a deep dive into the statistical properties of the dataset to understand types of questions and answer formats, tailoring the model's architecture accordingly.
• Feature Selection and Engineering: Identified key lexical and syntactic features that could enhance the LLM’s capabilities in question-answering tasks.
• Identifying Gaps and Biases: Carried out an examination of data to reveal inherent biases and limitations, laying the groundwork for model fine-tuning and ethical considerations.
• Model Evaluation Metrics: Set up appropriate evaluation metrics like F1-score and BLEU score based on a thorough understanding of the dataset, allowing for quantitative performance measurement.
• Correlation Analysis: Utilized Pearson correlation coefficients to understand interconnections between various data elements.

Algorithm Development:

• Finetuned a GPT-2 model for question answering: Employed transfer learning techniques to adapt a pre-trained GPT-2 model specifically for the question-answering domain, benefitting from the feature engineering and bias identification steps performed earlier.
• Developed an answer generation algorithm using the finetuned GPT-2 model: Created a custom algorithm to generate accurate and coherent answers from the finetuned model, demonstrating the potential for real-world applications like customer support automation.

• A/B Testing: Conducted A/B tests post-initial model training to quantitatively measure the model's real-world effectiveness.

Developed an AI-driven image compression solution by adapting the Transformer Neural Network architecture. Achieved a high compression rate of 0.5 bits/byte and 98.7% recovery accuracy, demonstrating the system's robustness and efficiency.

• Reviewed the statistical properties of image datasets to identify trends in color distribution and spatial characteristics.
• Performed feature engineering to isolated critical features, such as texture patterns and gradients, to improve the compression algorithm.
• Analyzed noise to understand its impact on compression and recovery accuracy.
• Enriched the training dataset using techniques like rotation, flipping, and scaling to ensure model robustness.

• Modified the encoder architecture from Transformer Neural Networks to specialize it for image compression tasks.
• Designed and implemented a decoder module.
• Utilized the CIFAR-10 dataset of diverse images for model training.
• Conducted grid search and random search to fine-tune the model's hyperparameters for optimal performance.
• Crafted a specialized loss function, composed of SSIM and MAE, focused on optimizing reconstruction and similarity errors to ensure high fidelity in decompressed images.

• Utilized PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) for performance measurement.

Built a network intrusion detection system Generative Adversarial Networks, leveraging the UNSW-NB15 dataset. The model demonstrated high applicability in real-world cybersecurity scenarios.

Data Preprocessing:

• Utilized the UNSW-NB15 dataset, applying data cleaning techniques to remove anomalies and inconsistencies, laying the foundation for robust training.

Data Analysis:

• Feature Importance: Employed machine learning algorithms to rank the importance of features like packet rate, byte size, and protocol type in intrusion detection.
• Anomaly Detection: Conducted in-depth analysis to identify patterns characteristic of network intrusions.
• Evaluation Metrics: Used metrics like True Positive Rate, False Positive Rate, and F1-score for quantitative performance assessment.
• Data Imbalance: Analyzed the class distribution to recognize and address data imbalance issues, which are common in cybersecurity datasets.
• Employed PCA (Principal Component Analysis) to reduce dimensionality and improve computation speed.

Model Engineering and Training:

• Trained Generative Adversarial Network model to build the intrusion detection algorithm, focusing on distinguishing between normal and malicious network traffic.

Testing:

• Conducted simulated attacks to test the system’s performance in realistic conditions.