Pritish92/lavida-oracle-stage2

该数据集是LaViDA项目的Oracle Stage-2部分的可重用数据，专注于数学推理。数据集包含了由Qwen2.5-Math-72B-Instruct生成的机器生成内容，并经过严格的过滤和处理。主要包含两种数据池：Oracle-augmented v2池（8,963个自我跟踪和3,354个过滤后的Oracle跟踪）和Oracle-only池（仅3,354个过滤后的Oracle跟踪）。数据集旨在用于重现LaViDA Oracle-augmented v2池，研究更强的教师跟踪如何改变7B潜在专家分布，以及训练或评估OracleAug变体。

This dataset repository contains the reusable Oracle Stage-2 artifacts for LaViDA: Latent Visitation Distribution Alignment for Mathematical Reasoning. The dataset includes machine-generated content by Qwen2.5-Math-72B-Instruct, filtered and processed through several steps. It features two types of pools: Oracle-augmented v2 pool (8,963 self traces + 3,354 filtered Oracle traces) and Oracle-only pool (3,354 filtered Oracle traces only). The dataset is intended for reproducing the LaViDA Oracle-augmented v2 pool, studying how stronger-teacher traces change the 7B latent expert distribution, and training or evaluating OracleAug variants.