Mattimax/mobile-actions-ita

---
license: cc-by-4.0
tags:
- gemma3
- gemma
- google
- functiongemma
- mobile-actions
- function-calling
- DAC
- M.INC.
language:
- it
size_categories:
- 10M<n<100M
task_categories:
- text-generation
---

## Dataset Card: Mobile Actions (Italian Adaptation for Function Calling)

### Overview

This dataset is an Italian adaptation of the original Google **Mobile Actions** dataset, designed to train lightweight models for on-device function calling. It preserves the original tool-calling schema in English while translating user interactions and contextual instructions into Italian.

The goal is to enable models to map natural language instructions in Italian to structured function calls, maintaining compatibility with existing tool interfaces defined in English.

This dataset is particularly suited for small, efficient models such as FunctionGemma 270M, where strict formatting and linguistic clarity are critical for performance.

---

### Motivation

The original Mobile Actions dataset focuses on English-language interactions. However, real-world deployment—especially on mobile devices—requires support for multilingual user inputs.

This adaptation addresses:

* Italian-language user queries
* Alignment with existing tool APIs (kept in English)
* Improved usability for Italian-speaking users
* Preservation of function-calling reliability in constrained models

---

### Key Characteristics

* **Language**:

* User messages: Italian
* Tool names: English (unchanged)
* Tool descriptions: Italian (adapted)
* **Task Type**:

* Semantic parsing → function calling
* **Domain**:

* Mobile OS actions (calendar, email, contacts, settings, maps)
* **Format**:

* JSONL (one sample per line)

---

### Dataset Structure

Each entry in the dataset is a JSON object with the following fields:

#### `metadata`

Indicates dataset split:

* `"train"`
* `"eval"`

#### `tools`

A list of available functions the model can call. Each tool includes:

* `name`: Function name (English, unchanged)
* `description`: Description of the function (Italian)
* `parameters`: JSON Schema-like structure defining inputs

#### `messages`

A sequence of conversational turns:

* `role`: `"developer"`, `"user"`, or `"assistant"`
* `content`: Natural language input (Italian)
* `tool_calls`: Expected structured function call (assistant only)

---

### Example

```json
{
"metadata": "train",
"tools": [
{
"function": {
"name": "create_calendar_event",
"description": "Crea un nuovo evento del calendario.",
"parameters": {
"type": "object",
"properties": {
"datetime": {
"type": "string",
"description": "La data e l'ora dell'evento nel formato AAAA-MM-GGTHH:MM:SS."
},
"title": {
"type": "string",
"description": "Il titolo dell'evento."
}
},
"required": ["title", "datetime"]
}
}
}
],
"messages": [
{
"role": "developer",
"content": "Data e ora correnti: 2025-06-04T15:29:23. Sei un modello che può eseguire chiamate di funzioni."
},
{
"role": "user",
"content": "Imposta un promemoria per una riunione di sincronizzazione del team venerdì alle 14."
},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "create_calendar_event",
"arguments": {
"title": "Riunione di sincronizzazione del team",
"datetime": "2025-06-06T14:00:00"
}
}
}
]
}
]
}
```

---

### Differences from the Original Dataset

Compared to the original Mobile Actions dataset by Google:

* User inputs are translated into Italian
* Tool descriptions are localized into Italian
* Function names and schema remain unchanged (English)
* Conversational style adapted to natural Italian phrasing
* Temporal expressions interpreted and normalized into ISO format

---

### Intended Use

This dataset is designed for:

* Fine-tuning small language models for function calling
* On-device AI assistants
* Multilingual semantic parsing research
* Mobile automation systems

---

### Limitations

* Limited to Italian language adaptation
* Assumes correct interpretation of temporal expressions
* Does not cover all possible Android system actions
* May inherit biases from the original dataset

---

### License

This dataset is derived from the original Mobile Actions dataset by Google. Users must comply with the licensing terms of the original dataset as specified on its Hugging Face page:
[https://huggingface.co/datasets/google/mobile-actions](https://huggingface.co/datasets/google/mobile-actions)

---

### Citation

If you use this dataset, please cite both the original dataset and this adaptation:

```
@misc{mobile_actions_google,
title = {Mobile Actions Dataset},
author = {Google},
year = {2024},
howpublished = {\url{https://huggingface.co/datasets/google/mobile-actions}}
}
```

```
@misc{mobile_actions_it,
title = {Mobile Actions Italian Adaptation for Function Calling},
author = {Your Name},
year = {2026},
note = {Italian adaptation of the Google Mobile Actions dataset}
}
```

---

### Acknowledgements

This work is based on the Mobile Actions dataset developed by Google, adapted to support Italian-language interactions while preserving compatibility with existing function-calling pipelines.