Korea-MES/Mixtral-Upperbound-V5

这是`Korea-MES/Mixtral-Upperbound-V4`数据集的改进版本，具有**更严格的MLT（最大长度令牌）范围**和**平衡的测试集**，以便在训练和评估期间更好地控制令牌。原始V4数据集存在两个主要问题：1. MLT范围过于宽泛，导致训练时混淆；2. 测试集不平衡，MLT分布不均。V5版本使用基于`Qwen/Qwen2.5-0.5B-Instruct`分词器的实际令牌计数的**紧密、精确的MLT范围**，以及**平衡的测试集**（每个MLT200个样本，按来源分层）。数据集包含多种MLT标签，每个标签有明确的令牌范围，并提供了详细的样本分布统计信息。

This is a refined version of `Korea-MES/Mixtral-Upperbound-V4` with **tightened MLT (Maximum Length Token) ranges** and a **balanced test set** for better token control during training and evaluation. The original V4 dataset had two main issues: 1. Overly broad MLT ranges, causing confusion during training; 2. Imbalanced test set with uneven MLT distribution. V5 uses **tight, precise MLT ranges** based on actual token counts using `Qwen/Qwen2.5-0.5B-Instruct` tokenizer and a **balanced test set** with 200 samples per MLT, stratified by source. The dataset includes various MLT labels with precise token ranges and provides detailed statistics on sample distribution.