Large Language Models with Religious Text (Undergraduate Thesis)• Fine-tuned Llama models using Low-Rank Adaptation (LoRA) to create a verse extraction tool for religious texts, employing large language models (LLMs) to answer user queries with relevant Bible verses​• Developed five custom NLP tasks to assess chatbot effectiveness in religious dialogues, including semantic similarity, cross-referencing, named entity recognition, and theological question answering​• Implemented web scrapers to gather and structure large datasets from Bible versions and religious sources, leading to enhanced chatbot performance across multiple tasks compared to ChatGPT• Evaluated model performance using exact match accuracy, cosine similarity, and clustering analysis to highlight theological NLP model capabilities and limitationsGender Bias in Machine Translation for Multilingual Masked Language Models • Addressed existing biases in previous methodologies and enhanced accuracy by removing confounding biases • Produced gendered lexicon to detect gendered sentences of different languages without parallel corporaAdvisor: Dr. Jinho Choi

Language Information and Computation Lab (LINC)• Explored innovative approaches to creating latently inseparable backdoor attacks in neural networks, particularly for text-based models, addressing the limitations of latent separability assumptions in existing defense methods• Designed and implemented backdoor attack mechanisms using partial and asymmetric triggers, enhancing attack stealth and robustness against latent-space-based defense methods• Analyzed and documented experimental results using attack success rates (ASR), perplexity, grammar error rate, and clustering metrics, contributing insights for future research on backdoor attacks in NLP• Prepared and presented research findings, contributing to a broader understanding of vulnerabilities in large language models and informing future work on adversarial attacks in NLPAdvisor: Dr. Jugal Kalita