CCCCCyx/20260526_repro_pkg

该数据集是一个轨迹数据集合，用于分析和比较不同优化器在训练过程中的性能。它通过计算每个训练检查点（如迭代步骤）的多个指标来生成轨迹，包括迭代次数(iter)、有效余弦相似度(cos_eff)、沿共享方向能量(E_phi)、垂直方向能量(E_perp)和共享方向能量比例(p_phi)等。这些指标基于模型参数的变化（如tau_t,p = theta_t,p - theta_0,p）和投影到预定义共享方向(phi_hat)上计算得出。数据集旨在重现原始研究中的轨迹数据（例如report_full2/data/trajectory_dir_energy.csv），但适用于新的优化器检查点序列，支持在任何具有Python和PyTorch的机器上使用。数据集中包含6个代表性参数（覆盖MLP下投影和注意力块）的聚合结果，用于高效分析优化器的收敛行为、能量分布和潜在问题（如能量扩散或崩溃）。This dataset is a collection of trajectory data used to analyze and compare the performance of different optimizers during training. It generates trajectories by computing multiple metrics for each training checkpoint (e.g., iteration step), including iteration (iter), effective cosine similarity (cos_eff), energy along the shared direction (E_phi), perpendicular energy (E_perp), and fraction of energy on the shared direction (p_phi). These metrics are derived from changes in model parameters (e.g., tau_t,p = theta_t,p - theta_0,p) and projections onto a predefined shared direction (phi_hat). The dataset is designed to reproduce trajectory data from the original study (e.g., report_full2/data/trajectory_dir_energy.csv) but for new optimizer checkpoint sequences, enabling use on any machine with Python and PyTorch. It aggregates results over 6 representative parameters (covering MLP down-projection and attention blocks) to efficiently analyze optimizer convergence behavior, energy distribution, and potential issues such as energy diffusion or collapse.