W-61/llama-multi-beta-testing-margin_dpo-hh-helpful-beta-0p01-20260429-085449-margin

该数据集是从一个Margin-DPO训练运行中导出的每一步的边缘摘要统计。数据集包含多个特征，如epoch、step、batch_size、mean、std、min、p10、median、p90、max、pos_frac、sample（每个示例的边缘）和npy（可选路径）。数据集来源于一个特定的训练运行，基础模型为`llama-3-8b-base-sft-hh-helpful-4xh200`，训练参数包括beta值为0.01，f_divergence_type为`reverse_kl`，f_alpha_divergence_coef为1.0。数据集混合器使用了`Anthropic/hh-rlhf`数据集，比例为1.0。

Per-step margin summary statistics exported from a Margin-DPO training run. The dataset includes features such as epoch, step, batch_size, mean, std, min, p10, median, p90, max, pos_frac, sample (per-example margins), and npy (optional path). The dataset originates from a specific training run with the base model `llama-3-8b-base-sft-hh-helpful-4xh200`, training arguments including beta=0.01, f_divergence_type=`reverse_kl`, f_alpha_divergence_coef=1.0. The dataset mixer uses the `Anthropic/hh-rlhf` dataset with a ratio of 1.0.