ferjorosa/bnqmark-20
收藏资源简介:
BNqMark-20是一个用于评估大型语言模型(LLMs)在离散贝叶斯网络中进行精确概率推理的基准数据集。它通过提供完整的条件概率表(CPT)规范并要求模型回答条件概率查询,将概率计算与语言解释隔离开来。数据集包括78个贝叶斯网络(4-20个二元变量)、434个条件概率查询和来自9个前沿模型的7,812个LLM评估结果。实验涵盖两种协议:原始推理(模型直接从CPT计算概率)和代码生成(模型编写Python代码解决相同的推理问题)。数据集结构分为三个配置:贝叶斯网络配置(bns)、推理查询(queries)和LLM评估结果(experiments)。
BNqMark-20 is a benchmark dataset for evaluating Large Language Models (LLMs) on exact probabilistic inference in discrete Bayesian Networks. It isolates probabilistic computation from linguistic interpretation by giving models complete conditional probability table (CPT) specifications and asking them to answer conditional probability queries. The dataset includes 78 Bayesian networks with 4-20 binary variables, 434 conditional probability queries, and 7,812 LLM evaluation results from 9 frontier models. Experiments cover two protocols: raw reasoning, where the model computes the probability directly from the CPTs, and code generation, where the model writes Python code to solve the same inference problem. The dataset is organized into three separate configurations: Bayesian Network Configurations (bns), Inference Queries (queries), and LLM Evaluation Results (experiments).




