aicrowd/arc-whestbench-public-2026
收藏资源简介:
WhestBench 2026: ARC白盒估计挑战数据集是一个用于白盒激活估计的基准测试。它包含预构建的随机初始化ReLU多层感知机(MLP)及其真实激活统计信息。给定MLP的权重和严格的浮点操作预算,目标是预测每个神经元在标准高斯输入下的平均后激活值。数据集提供两个独立的分割:mini(100个MLP,用于开发迭代)和full(1000个MLP,用于正式评估),它们使用不同的种子生成,互不重叠。每个MLP包含权重矩阵、各层激活均值、最终层激活均值等列,真实值通过蒙特卡洛模拟(10亿样本)计算得出,具有高精度。该数据集旨在支持机器学习模型在有限计算预算下对神经网络内部激活的估计能力评估。
WhestBench 2026: ARC White-Box Estimation Challenge Dataset is a benchmark for white-box activation estimation. It contains pre-built randomly initialized ReLU multilayer perceptrons (MLPs) paired with their ground-truth activation statistics. Given the weights of an MLP and a strict floating-point operation budget, the objective is to predict the average post-activation values of each neuron when the input follows a standard Gaussian distribution. The dataset provides two independent, non-overlapping splits: the mini split (100 MLPs for iterative development) and the full split (1000 MLPs for formal evaluation), which are generated using different random seeds. Each MLP includes columns such as weight matrices, activation means of each layer, activation means of the final layer, and other relevant columns. The ground-truth values are calculated via Monte Carlo simulation with 1 billion samples and feature high accuracy. This dataset aims to support the evaluation of machine learning models' ability to estimate internal neural network activations under limited computational budgets.




