Nlp

LLM-based applications face security challenges in form of prompt injections and jailbreaks. This project reviews the key architectural improvements underpinning ModernBERT, and implements fine-tuning for discriminating malicious prompts. PangolinGuard closely approximates the performance of Claude 3.7 on a mixed benchmark, while maintaining low latency (< 40ms).

The ever-growing volume of research publications necessitates efficient methods for structuring such knowledge. This automated solution uses Machine Learning (UMAP, HDBSCAN), Embedding Quantization, and an LLM pipeline to classify 25,000 arXiv publications under a novel taxonomy.