W-61/llama-3-8b-base-new-dpo-hh-harmless-4xh200-batch-64-q_t-0.45-s_star-0.4-eta-0.5-margin-log

该数据集是从一个New-DPO训练运行中导出的每一步的边缘统计摘要。包含训练过程中的多个统计特征，如epoch（训练轮次）、step（步骤）、batch_size（批次大小）、mean（平均值）、std（标准差）、min（最小值）、p10（第10百分位数）、median（中位数）、p90（第90百分位数）、max（最大值）、pos_frac（正分数）、sample（每个步骤的有效批次的边缘值样本）和npy（可选，保存完整边缘数组的路径）。数据集来源于特定的模型训练运行，使用了Anthropic/hh-rlhf数据集进行混合训练。This dataset is a per-step marginal statistical summary derived from a New-DPO training run. It includes multiple statistical features collected during the training process, such as epoch (training epoch), step (training step), batch_size (batch size), mean (mean value), std (standard deviation), min (minimum value), p10 (10th percentile), median (median value), p90 (90th percentile), max (maximum value), pos_frac (positive fraction), sample (samples of marginal values from valid batches per step), and npy (optional, path to save the complete marginal array). This dataset originates from a specific model training run that utilized the Anthropic/hh-rlhf dataset for mixed training.