vigneshwar234/phantasm-hallucination-benchmark
收藏资源简介:
PHANTASM幻觉基准数据集是PHANTASM框架的配套数据集,这是首个通过数学方法将大型语言模型(LLM)的幻觉、虚构和认知校准错误转化为生产性特征的机器学习系统。该数据集支持评估PHANTASM的三个核心支柱:HGT(幻觉边界检测)、CMN(虚构新颖性/合理性评分)和UC(不确定性校准)。具体包括以下分割:hgt_train(30个示例用于幻觉边界检测训练)、hgt_test(10个示例用于测试)、cmn_train(10个示例用于虚构新颖性和合理性评分)和uc_train(20个示例用于不确定性校准)。数据集涵盖多个领域(如历史、科学、医学),并提供标签如幻觉标签、新颖性分数、合理性分数和校准置信度等,旨在帮助研究者和开发者评估和改进LLM在生成文本时的可靠性、准确性和不确定性管理。
The PHANTASM Hallucination Benchmark is the companion dataset to the PHANTASM framework — the first ML system to mathematically invert LLM hallucination, confabulation, and epistemic miscalibration into productive features. This dataset supports evaluation of all three PHANTASM pillars: HGT (Hallucination boundary detection), CMN (Confabulation novelty/plausibility scoring), and UC (Uncertainty calibration). It includes splits such as hgt_train (30 examples for hallucination boundary detection training), hgt_test (10 examples for testing), cmn_train (10 examples for confabulation novelty and plausibility scoring), and uc_train (20 examples for uncertainty calibration). The dataset covers various domains (e.g., history, science, medicine) and provides labels like hallucination_label, novelty_score, plausibility_score, and calibrated_confidence, aiming to assist researchers and developers in evaluating and improving LLM reliability, accuracy, and uncertainty management in text generation.




