Finetuning with Scientific Data Increases Hallucinations: A Multi-domain Factuality Evaluation of LLMs

The paper introduces SciFactCheck, a benchmark consisting of 2,500 prompts across five scientific domains, designed to evaluate factuality hallucinations in large language models (LLMs). The study assesses 18 LLMs, revealing that scientifically fine-tuned models exhibit degraded factual reliability and increased linguistic assertiveness despite lower internal confidence. This research highlights the risks of hallucinations in fine-tuned models for scientific applications and emphasizes the need for enhanced verification methods for scientific content.