fnbm-current-sweep-results
收藏资源简介:
该数据集名为fnbm-current-sweep-results,是专为fnbm-current模型设计的网格参数扫描试验结果集合。它包含59个试验样本,每个样本对应一次参数组合的模型训练与评估运行,共计29个字段。核心字段包括试验唯一标识符(trial_id)、实验名称(experiment_name)、扫描名称(sweep_name)、参数组合名称(combo_name)、扫描参数覆盖配置(grid_overrides)、完整配置YAML(config_yaml)、集群作业信息(job_id, cluster)和记录时间戳(timestamp)。性能评估指标涵盖最佳验证损失(best_val_loss)及其对应的皮尔逊相关系数(best_val_pearson_r)和轮次(best_epoch)、仿真数据的最终轮次R²(sim_r2)和皮尔逊相关系数(sim_pearson_r)、一阶与二阶贡献比例分析指标(如sim_mean_gt_z_abs_ratio, sim_mean_model_z_abs_ratio, absorption_gap)、滤波器激活分散度指标(promiscuity_mean_entropy, promiscuity_max_entropy)、真实交互对数量(n_true_interaction_pairs)与模型有效交互对数量(interaction_effective_num_pairs)及其差异(interaction_concentration_gap)。此外,还包含每轮历史指标(epoch_history)、Weights & Biases运行链接(wandb_run_url)和本地报告路径(report_pdf_path)。该数据集适用于机器学习模型超参数调优分析、模型性能对比、训练动态研究以及可解释性分析(如贡献分解)等任务。
The dataset is named fnbm-current-sweep-results and is a collection of grid parameter sweep experiment results specifically for the fnbm-current model. It contains 59 trial samples, each corresponding to a model training and evaluation run with a parameter combination, totaling 29 fields. Key fields include the trial unique identifier (trial_id), experiment name (experiment_name), sweep name (sweep_name), parameter combination name (combo_name), grid parameter overrides configuration (grid_overrides), full configuration YAML (config_yaml), cluster job information (job_id, cluster), and timestamp (timestamp). Performance evaluation metrics cover the best validation loss (best_val_loss) and its corresponding Pearson correlation coefficient (best_val_pearson_r) and epoch (best_epoch), final epoch R² (sim_r2) and Pearson correlation coefficient (sim_pearson_r) for simulation data, first-order and second-order contribution ratio analysis metrics (e.g., sim_mean_gt_z_abs_ratio, sim_mean_model_z_abs_ratio, absorption_gap), filter activation dispersion metrics (promiscuity_mean_entropy, promiscuity_max_entropy), number of true interaction pairs (n_true_interaction_pairs) versus model effective interaction pairs (interaction_effective_num_pairs) and their difference (interaction_concentration_gap). Additionally, it includes per-epoch history metrics (epoch_history), Weights & Biases run links (wandb_run_url), and local report paths (report_pdf_path). This dataset is suitable for tasks such as machine learning model hyperparameter tuning analysis, model performance comparison, training dynamics research, and interpretability analysis (e.g., contribution decomposition).
数据集概述:fnbm-current-sweep-results
该数据集记录了 fnbm-current 模型的网格扫描(Grid-sweep)试验结果。
- 许可协议:MIT
- 数据量:包含 71 行记录,29 列字段。
- 数据用途:存储每次试验的性能指标、模拟对比结果以及配置信息,用于分析模型在不同参数组合下的表现。
字段说明
数据集包含以下关键字段,分为几类:
试验标识与配置:
trial_id:试验的唯一标识符。experiment_name:试验所属的实验文件夹名称。sweep_name:整个扫描过程的基运行名称。combo_name:基于基名称与参数值生成的单次运行名称。grid_overrides:JSON编码的扫描参数键值对。config_yaml:本次试验的完整解析配置。job_id/cluster:集群任务ID和集群名称。
模型性能:
best_val_loss:训练过程中达到的最佳验证损失。best_val_pearson_r:最佳验证损失时期的验证皮尔逊相关系数。best_epoch:最佳验证损失所在的训练轮次。
模拟对比指标:
sim_r2:最终训练轮次时,模型预测与模拟真值之间的R²。sim_pearson_r:最终训练轮次时,与模拟真值的皮尔逊相关系数。sim_mean_gt_z_abs_ratio/sim_mean_model_z_abs_ratio:真值与模型预测中二阶效应贡献的平均比例。sim_z_abs_1_corr/sim_z_abs_2_corr:模型与真值在样本级一阶和二阶贡献上的相关性。absorption_gap:模型高估二阶效应的程度指标。
其他指标:
promiscuity_mean_entropy/promiscuity_max_entropy:滤波器激活的漫射程度代理指标。n_true_interaction_pairs/interaction_effective_num_pairs:模拟中真实与有效的相互作用对数量。epoch_history:JSON格式的逐轮次指标历史。wandb_run_url:W&B运行页面链接。report_pdf_path:试验报告PDF文件的路径。
使用方式
可通过 Hugging Face datasets 库加载该数据集:
python from datasets import load_dataset
dataset = load_dataset("arushram/fnbm-current-sweep-results", split="train") print(f"Loaded {len(dataset)} rows")




