Hallucination and Bias mitigation tool for LLM AI models: Extractor v1.1

The project addresses critical shortcomings in large language models (LLMs), particularly hallucinations (factually incorrect yet plausible-sounding outputs), embedded biases, and the inclusion of copyrighted material in training data. These issues compromise the trustworthiness, fairness, and legal compliance of AI systems. Extractor is a novel tool developed by Multiverse Computing that enables precise identification and removal of problematic data—whether biased, hallucination-inducing, or copyrighted—without retraining the model or altering its core architecture. 

Using cutting-edge model editing techniques (MEMIT and PMET), Extractor empowers developers to "make the model forget" selected content, increasing output reliability while supporting compliance with ethical, legal, and societal norms. HPC access through EuroHPC's Leonardo BOOSTER enables the efficient dual-model benchmarking (pre- and post-removal) necessary to validate this approach on resource-intensive LLMs like LLaMA3.1-8B and Mixtral8-7B.