ellabettison/reasoning-distillation-data

推理蒸馏数据集是一个通过推理蒸馏过程生成的数据集，主要用于从ER管道分类文件中提取（文档，主题配置文件）对，并使用前沿模型为每个输入生成多个推理轨迹。通过自一致性验证后，数据集还通过对抗性DPO拒绝策略进行增强。数据集包含多个配置文件，如sft（用于监督微调）、dpo（用于直接偏好优化）、eval（用于评估模型在棘手案例上的表现）等。

The reasoning distillation dataset is generated through a reasoning distillation process, primarily aimed at extracting (document, subject profile) pairs from ER pipeline triage files and generating multiple reasoning traces per input using a frontier model. After verification via self-consistency, the dataset is further augmented with adversarial DPO rejection strategies. The dataset includes multiple configurations such as sft (for supervised fine-tuning), dpo (for direct preference optimization), eval (for evaluating model performance on tricky cases), etc.