almanach/topxgen-llama-4-scout-CompTra

该数据集是基于大型语言模型进行机器翻译推理的合成数据集。它通过在翻译前生成中间推理标记来增强模型的能力。数据集的每条记录包含源句子、目标翻译、源语言、目标语言、源句子分解的短语及其翻译、选择的高质量翻译以及用于CoT fine-tuning的目标序列。数据集支持Xhosa语言，并用于微调LLM以进行基于推理的翻译、研究MT中的思考标记以及比较CoTFT和IOFT在MT设置中的效果。

This dataset is a synthetic dataset for reasoning-based machine translation using large language models. It enhances model capabilities by generating intermediate reasoning tokens before translation. Each record in the dataset contains the source sentence, target translation, source language, target language, phrases obtained by decomposing the source sentence and their translations, the selected high-quality translation, and the target sequence for CoT fine-tuning. The dataset supports the Xhosa language and is used for fine-tuning LLMs for reasoning-based translation, studying thinking tokens in MT, and comparing CoTFT vs IOFT in MT setups.