jackf857/llama-3-8b-base-new-dpo-ultrafeedback-4xh200-batch-128-q_t-0.43-s_star-0.4-margin

该数据集是从New-DPO训练运行中导出的每步边际摘要统计数据集。来源运行信息包括模型仓库ID、基础模型、训练运行名称、W&B项目、训练器类型等。边际训练参数包括beta、f_divergence_type、f_alpha_divergence_coef、s_star、eta和q_t等。数据集包含的列有epoch、step、batch_size、mean、std、min、p10、median、p90、max、pos_frac、sample（每步有效批次的每示例边际）和npy（保存完整边际数组的可选路径）。数据集混合器使用了HuggingFaceH4/ultrafeedback_binarized。

Per-step margin summary statistics exported from a New-DPO training run. Source run information includes model repo id, base model, training run name, W&B project, trainer type, etc. Margin training arguments include beta, f_divergence_type, f_alpha_divergence_coef, s_star, eta, and q_t. Columns include epoch, step, batch_size, mean, std, min, p10, median, p90, max, pos_frac, sample (per-example margins for the effective batch on that logged step), and npy (optional path to the saved full margin array). Dataset mixer uses HuggingFaceH4/ultrafeedback_binarized.