Demo: Statistically Significant Results on Biases and Errors of LLMs Do Not Guarantee Generalizable Results

Here we present our generated datasets for the paper Demo: Statistically Significant Results On Biases and Errors of LLMs Do Not Guarantee Generalizable Results. We generated our (3.2M prompts, 29K answers, and 684K evaluations) dataset as follows: First, we generated our prompts from 37 disorders with 68 symptoms, 79 desires, 518 distinct patients, and 48 styles using Llama3-ChatQA-1.5-8B with a temperature of 0.1. Next, we generated 29,256 answers for a subset of the questions using three LLMs—Llama3-ChatQA-1.5-8B, BioMistral-7B, and MedGemma-4B-it—each with a temperature of 0.1. Finally, for each contrastive set of answers (corresponding to prompts that differ only in terms of the Patient Expression), we applied a filter based on semantic embeddings to remove answers that are too similar (similarity threshold: 0.7). We then applied our hallucination-detection, omission-detection, and treatment-evaluation pipelines using three LLMs (Llama3-ChatQA-1.5-8B, Qwen2.5-7B-Instruct, and OLMo-2-1124-13B) at a temperature of 0.1. We additionally evaluated each set of answers with our agentic systems using Mistral’s Nemo-Instruct-2407 model at 0 temperature. We're providing commands we ran, generated questions, answers, detections, and plotting script with this release. Acknowledgments This research was, in part, funded by the Advanced Research Projects Agency for Health (ARPA-H). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Government.