mobile-actions

# Mobile Actions: A Dataset for On-Device Function Calling

The dataset contains conversational traces designed to train lightweight models (such as FunctionGemma 270M) to translate natural language instructions into executable function calls for Android OS system tools.

## Dataset Format

The dataset is provided in JSONL format. Each line represents a data sample. The
dataset is pre-split into training and evaluation sets. This distinction is
denoted by the `metadata` field within each sample.

Each JSON object in the file contains the following fields:

- `metadata`: Contains metadata about the data sample for splitting. The value is either "train" or "eval".
- `tools`: A list of available tools (functions) that the model can call. Each tool has:
- `function`: An object describing the function:
- `name`: The name of the function.
- `description`: A description of what the function does.
- `parameters`: An object describing the parameters the function accepts, following a JSON Schema like structure.
- `messages`: A list of messages, usually containing user input and the expected function call.
- `role`: Typically "user" for the input command and "assistant" for the function call.
- `content`: The natural language input from the user.
- `tool_calls`: (For assistant role) A list of tool calls the model should predict. Each tool call has:
- `id`: A unique identifier for the tool call.
- `type`: Always "function".
- `function`: An object specifying the function to call:
- `name`: The name of the function.
- `arguments`: A stringified JSON object containing the arguments for the function.

# 移动操作：面向设备端函数调用的数据集

本数据集包含对话轨迹数据，旨在训练轻量级模型（如FunctionGemma 270M），将自然语言指令转换为适用于安卓（Android）操作系统系统工具的可执行函数调用。

## 数据集格式

本数据集采用JSONL格式存储，每一行对应一条数据样本。数据集已预先划分为训练集与评估集，该划分信息通过每条样本中的`metadata`字段进行标注。

文件中的每个JSON对象包含以下字段：
- `metadata`：用于数据划分的样本元数据，取值仅为"train"或"eval"，分别对应训练集与评估集。
- `tools`：模型可调用的可用工具（函数）列表。每个工具包含如下结构：
- `function`：描述函数的对象，包含：
- `name`：函数的名称
- `description`：函数的功能说明
- `parameters`：描述函数接收参数的对象，采用类似JSON Schema的结构。
- `messages`：消息列表，通常包含用户输入与预期的函数调用结果：
- `role`：消息角色，通常"user"代表用户输入的指令，"assistant"代表需生成的函数调用。
- `content`：用户提供的自然语言输入内容。
- `tool_calls`：（仅适用于assistant角色）模型应预测的工具调用列表。每个工具调用包含：
- `id`：工具调用的唯一标识符
- `type`：固定取值为"function"
- `function`：指定待调用函数的对象，包含：
- `name`：函数名称
- `arguments`：字符串化的JSON对象，存储函数所需的调用参数。