sardinelab/DocBlocks

DocBlocks是一个高质量的多语种文档级机器翻译数据集，旨在用于长上下文翻译任务的大型语言模型（LLM）微调。与传统基于句子的数据集不同，它包含完整的文档，具有自然的语篇结构和上下文对齐，帮助模型在较长的文本中保持连贯性、一致性和高翻译质量。

DocBlocks is a high-quality, multilingual document-level machine translation (MT) dataset designed to fine-tune large language models (LLMs) on long-context translation tasks. Unlike traditional sentence-level datasets, it contains full documents with natural discourse structures and contextual alignment, helping models maintain coherence, consistency, and high translation quality across longer texts.