emre/lima_dirty_tr

LIMA土耳其语翻译与对齐数据集是一个基于LIMA研究的土耳其语翻译和特别构建的数据集，专为对齐技术设计。它不仅包括翻译，还包括用于直接与流行的基于偏好的学习算法（如DPO、ORPO、PPO和SimPO）一起使用的`chosen`和`rejected`响应对。数据集通过为每个原始示例生成三个版本的响应（原始翻译、更好、更差），从而形成了三个偏好对，以此丰富数据集。这允许模型学习生成更细微和更高质量的土耳其语输出。

The Lima Turkish Translated & Engineered for Alignment Dataset is a Turkish translation and specially structured version of the original LIMA dataset, inspired by the LIMA (Less Is More for Alignment) study. It is engineered specifically for alignment techniques and includes `chosen` and `rejected` response pairs designed for direct use with popular preference-based learning algorithms such as DPO (Direct Preference Optimization), ORPO (Odds Ratio Preference Optimization), PPO (Proximal Policy Optimization), and SimPO (Simple Preference Optimization). The dataset enriches itself by generating three versions of responses (original translation, better, worse) for each original example, forming three preference pairs. This allows models to learn to generate more nuanced and higher-quality Turkish outputs.