home_assistant_train_ru
收藏资源简介:
Home Assistant Train RU 是 acon96/Home-Assistant-Requests-V2 数据集的俄语翻译版本,包含 41,798 个示例,专为俄语小型 Home Assistant 模型的微调设计,支持工具调用、意图分类和设备控制等任务。数据集对用户查询和助手面向客户的自然语言文本回复进行了完整俄语翻译(采用非正式的“ты”称呼形式),而系统提示、工具描述、工具调用以及设备状态信息(例如“blinds.kitchen is now shut.”)则保留原始英文,因为这些内容由 Home Assistant 系统内部解析,不应更改。翻译工作使用 nvidia/nemotron-3-ultra-550b-a55b 模型完成,采用每批 50 行、5 个并行会话的处理方式,并完全保留了原始的 JSONL 结构和 messages 字段格式。每条数据记录是一个 JSON 对象,其 messages 字段遵循 OpenAI 聊天格式,其中 content 字段是由 type 和 text 组成的对象数组。数据集中特意未翻译设备状态和实体 ID(如 blinds.kitchen),在适配特定工具时需要单独的实体解析器进行处理。该数据集适用于文本生成任务,主要语言为俄语和英语,规模在 1 万到 10 万样本之间。
Home Assistant Train RU is a Russian translation of the acon96/Home-Assistant-Requests-V2 dataset, containing 41,798 examples, intended for fine-tuning small Home Assistant models in Russian, supporting tasks such as tool calling, intent classification, and device control. The dataset includes a full Russian translation of user queries and assistant-facing natural language text responses (using the informal ты form), while system prompts, tool descriptions, tool calls, and device status information (e.g., blinds.kitchen is now shut.) remain in their original English, as these are parsed internally by the Home Assistant system and should not be altered. Translation was performed using the nvidia/nemotron-3-ultra-550b-a55b model, with a batch size of 50 lines and 5 parallel sessions, fully preserving the original JSONL structure and messages field format. Each data record is a JSON object with a messages field following the OpenAI chat format, where the content field is an array of objects with type and text. Device states and entity IDs (e.g., blinds.kitchen) are intentionally left untranslated in the dataset, requiring separate entity parsers for adaptation to specific tools. The dataset is suitable for text generation tasks, primarily in Russian and English, with a scale between 10,000 and 100,000 samples.
数据集概述
- 数据集名称: Home Assistant Train RU
- 许可证: MIT
- 任务类型: 文本生成
- 语言: 俄语(ru)、英语(en)
- 数据集大小: 约41,798个样本,规模在10K到100K之间
描述
该数据集是 acon96/Home-Assistant-Requests-V2 的俄语化版本,专为在俄语环境下微调小型 Home Assistant 模型设计,涵盖工具调用、意图分类和设备管理等任务。
翻译内容
- User 请求: 完全翻译为俄语,使用非正式的“ты”形式。
- Assistant 文本回复: 自然语言回复部分已翻译。
- 未翻译内容: System 提示、工具描述、函数描述、工具调用及设备状态(如
blinds.kitchen is now shut.)保留英文,以确保 Home Assistant 解析不受影响。
翻译方法
采用 NVIDIA 的 nvidia/nemotron-3-ultra-550b-a55b 模型(API,关闭推理模式)进行翻译,批次50行,并行5个会话,原始 JSON 行结构(包含 messages 字段)完全保留。
数据结构
每条数据为一个 JSON 对象,包含 messages 字段,格式为 OpenAI 聊天格式。示例结构如下:
json { "messages": [ { "role": "system", "content": [ { "type": "text", "text": "..." } ] }, { "role": "user", "content": [ { "type": "text", "text": "включи свет на кухне" } ] }, { "role": "assistant", "content": [ { "type": "text", "text": "Сейчас включу свет на кухне." } ] }, { "role": "assistant", "content": [ { "type": "text", "text": "light.kitchen is now on." } ] }, { "role": "tool", "content": [ { "type": "text", "text": "..." } ] } ] }
使用方式
通过 Hugging Face datasets 库加载:
python from datasets import load_dataset ds = load_dataset("RockMan256/home_assistant_train_ru")
注意事项
设备状态和实体 ID(如 blinds.kitchen、light.kitchen)未被翻译,用户需在适配具体工具时通过单独的实体解析器处理。




