LLM-PIDS: Data, Model Weights and Evaluation Artefacts

Supporting artefacts for the paper LLM-PIDS: Prompt Classification for Detection and Mitigation of Prompt Injection Attacks. This record includes the preprocessed corpus, model weights trained on this corpus, and the results of evaluation of a controlled experiment comparing three approaches to prompt injection attack detection: traditional machine learning (based on TF-IDF and LinearSVC classifiers), deep learning (BiLSTM neural networks with GloVe embeddings), and transfer learning (DistilBERT and ModernBERT base). Source code, including corpus construction scripts and all training and evaluation notebooks, is available at https://github.com/ver1812/Capstone_Project Source datasets: deepset/prompt-injections, allenai/wildjailbreak, ahsanayub/malicious-prompts, and MAlmasabi/Indirect-Prompt-Injection-BIPIA-GPT, all available on Hugging Face.