Weichuan/TC-DN-MT-Genes

该数据集发布了用于研究温度约束非确定性机器翻译（ND-MT）与现代大型语言模型（LLMs）结合的原始机器翻译生成数据。每个JSON文件存储了每个源句子的一个或多个随机或贪婪假设，以及来自标准WMT风格测试集的参考翻译。数据集包含多种语言对的翻译生成，如英语-中文（EN-ZH）、英语-德语（EN-DE）和英语-俄语（EN-RU），并提供了贪婪解码和温度采样的多种翻译假设。数据以JSON格式存储，每个文件包含源句子、参考翻译和候选翻译列表。数据集旨在支持ND-MT分析、多假设MT指标的训练和温度及样本大小对翻译多样性和质量影响的研究。

This repository releases raw machine translation generations used in the study of temperature-constrained non-deterministic machine translation (ND-MT) with modern large language models (LLMs). Each JSON file stores, for every source segment, one or more stochastic or greedy hypotheses together with the reference translation from standard WMT-style test sets. The dataset includes translations for multiple language pairs such as English-Chinese (EN-ZH), English-German (EN-DE), and English-Russian (EN-RU), and provides various translation hypotheses through greedy decoding and temperature sampling. The data is stored in JSON format, with each file containing source sentences, reference translations, and lists of candidate translations. The dataset is intended for ND-MT analysis, training of multi-hypothesis MT metrics, and studying the effects of temperature and sample size on translation diversity and quality.