brandonbaek/konglish-synthetic-instruct
收藏资源简介:
Bori V2是一个合成生成的指令跟随数据集,专门设计用于教授小型语言模型(SLMs)的双语(韩语/英语)能力、自然语码转换(Konglish)和对话翻译。该数据集使用先进的商业大语言模型(DeepSeek V4)通过自指导范式构建,旨在高度清洁、多样且语言自然。其特点包括任务多样性(涵盖对话语码转换、直接翻译请求、语法映射、Konglish习语解释和双语逻辑推理),格式采用标准的Alpaca JSON格式(包括instruction、input、output字段),以确保与标准HuggingFace Trainer和SFTTrainer管道无缝兼容,并专注于现代对话语境(如办公环境、软件开发、日常社交对话),这些语境中英语-韩语语码转换在日常韩语交流中自然发生。数据集的每个样本包含instruction(提示或任务请求,可以是英语、韩语或混合的Konglish)、input(上下文背景,可选,默认为空字符串)和output(目标黄金标准响应,展示高度自然和有用的双语/Konglish生成)。该数据集旨在用于双语小型语言模型(如SmolLM-135M或Qwen2-0.5B)的监督微调(SFT),以及对话聊天机器人中语码转换能力的对齐和评估。
This is a synthetically generated instruction-following dataset designed specifically to teach bilingual (Korean/English) capabilities, natural code-switching (Konglish), and conversational translation to Small Language Models (SLMs). It was constructed using advanced commercial large language models (DeepSeek V4) utilizing self-instruct paradigms, designed to be highly clean, diverse, and linguistically natural. The dataset features task diversity including conversational code-switching, direct translation requests, grammatical mapping, explanation of Konglish idioms, and bilingual logical reasoning. It is structured in standard Alpaca JSON format (instruction, input, output) for seamless compatibility with standard HuggingFace Trainer and SFTTrainer pipelines, and is quality-centric, focused on modern conversational contexts (office environments, software development, daily social dialogue) where English-Korean code-switching naturally occurs in everyday Korean communication. Each sample contains instruction (the prompt or task request, can be in English, Korean, or mixed Konglish), input (contextual background, optional, empty string by default), and output (the target gold-standard response demonstrating highly natural and helpful bilingual/Konglish generation). The dataset is intended for Supervised Fine-Tuning (SFT) of bilingual Small Language Models (such as SmolLM-135M or Qwen2-0.5B) and alignment & evaluation of code-switching capabilities in conversational chatbots.





