---
license: gpl-3.0
---

# DiaVio Dataset

## Overview

Simulation testing has been widely adopted by leading companies to ensure the safety of autonomous driving systems (ADSs). A number of scenario-based testing approaches have been developed to generate diverse driving scenarios for simulation testing, and demonstrated to be capable of finding safety violations. However, there is no automated way to diagnose whether these violations are caused by the ADS under test and which category these violations belong to. As a result, great effort is required to manually diagnose violations.

To bridge this gap, we propose DiaVio to automatically diagnose safety violations in simulation testing by leveraging large language models (LLMs). It is built on top of a new domain specific language (DSL) of crash to align real-world accident reports described in natural language and violation scenarios in simulation testing. DiaVio fine-tunes a base LLM with real-world accident reports to learn diagnosis capability, and uses the fine-tuned LLM to diagnose violation scenarios in simulation testing. Our evaluation has demonstrated the effectiveness and efficiency of DiaVio in violation diagnosis.

## Dataset Contents

This dataset.zip comprises three main components:

### 1. Accident Reports

A collection of real-world accident reports described in natural language. These reports serve as the basis for fine-tuning the language model to enhance its diagnosis capability.

### 2. Extracted Information

Information extracted from the accident reports, including key details about the accidents. This data is used to train DiaVio and enable it to associate real-world accidents with simulation testing scenarios.

### 3. GroundTruth

The ground truth data, indicating the correct diagnosis for each violation scenario. This dataset helps evaluate the effectiveness and efficiency of DiaVio in diagnosing safety violations.

## How to Use

Researchers and practitioners in the field of autonomous driving systems can utilize this dataset to:

- Train and fine-tune language models for safety violation diagnosis.
- Evaluate and compare the performance of different diagnostic systems.
- Enhance the capabilities of automated diagnosis tools in simulation testing.

## Citation

If you use this dataset in your research or work, please cite our related publication.