Mobile Actions is a dataset released by Google in 2025 for on-device function calling. It is designed to train lightweight language models to map natural language operation instructions to executable Android system function calls to support mobile smart assistants and automated tasks.

The dataset is organized in the form of dialogue trajectories. Each sample contains user operation instructions and their corresponding function call annotations, and is divided into a training set (train) and an evaluation set (eval) based on metadata.

Data Structures:

Each JSON sample mainly contains the following:

• metadata: Sample partitioning information, used to distinguish between the training set (train) and the evaluation set (eval);
• tools: A list of system functions that the model can call in the current sample. Each function includes a function name, function description, and parameter definition.
• Messages: A sequence of dialogue messages, typically including natural language operation instructions given by the user (user) and the corresponding model output (assistant).
• tool_calls (appears only in assistant messages): The specific function call information that the model needs to predict, including the function name and its parameters, which are represented in JSON format.