aicrowd/whestbench-ci-fixture
收藏资源简介:
WhestBench是一个用于白盒激活估计的基准测试数据集。具体任务是在给定小型ReLU多层感知机(MLP)的权重和严格的浮点操作(FLOP)预算下,预测当网络输入标准高斯分布时每个神经元的平均后激活值。这是WhestBench 2026评估数据集,包含100个MLP,分为两个分割:公开分割(50个MLP)用于公共排行榜实时评分,保留分割(50个MLP)用于私有/最终排行榜,分数仅在轮次结束时公布。每个MLP数据包括权重、真实激活均值、最终层均值、平均方差等列,用于评估估计器的性能。数据集宽度为256,深度为8,每个MLP使用1,024个蒙特卡洛样本计算真实值。
WhestBench is a benchmark for white-box activation estimation: given the weights of a small ReLU multi-layer perceptron (MLP) and a strict floating-point-operation (FLOP) budget, predict the average post-activation value of every neuron when the network is fed standard Gaussian inputs. This is the WhestBench 2026 Evaluation Dataset, containing 100 MLPs with two disjoint splits: public split (50 MLPs) for real-time scoring on the public leaderboard, and holdout split (50 MLPs) for the private/final leaderboard revealed at the conclusion of the round. Each MLP includes columns such as weights, ground-truth activation means, final-layer means, average variance, etc., for evaluating estimator performance. The dataset has a width of 256, depth of 8, and uses 1,024 Monte Carlo samples per MLP for ground truth computation.



