johanneskirmayr/car-bench-dataset

CAR-Bench Dataset

CAR-Bench is a benchmark for evaluating AI voice assistants in a realistic automotive (car) environment. It tests an agent's ability to correctly use vehicle control tools, handle disambiguation, and avoid hallucinations.

Dataset Structure

The dataset is organized into task configs and mock data configs:

Tasks

Each task defines a user persona, an instruction, the initial vehicle/environment context, and the ground-truth sequence of tool-call actions the assistant should perform.

Config	Description	Train	Test
tasks_base	Standard tasks covering vehicle controls, navigation, calendar, etc.	50	50
tasks_disambiguation	Tasks requiring the agent to disambiguate parameters (internally via preferences or by asking the user)	30	26
tasks_hallucination	Tasks where certain tools/parameters are intentionally removed to test if the agent hallucinates	48	50

Task schema:

Field	Type	Description
task_id	string	Unique task identifier
persona	string	Description of the simulated user's personality and communication style
calendar_id	string	Reference to a calendar in the mock data
instruction	string	The instruction given to the simulated user
context_init_config	string (JSON)	Initial vehicle and environment state (battery, seats, location, weather, preferences, etc.)
actions	string (JSON)	Ground-truth sequence of tool calls [{name, kwargs, index, dependent_on_action_index}]
task_type	string	One of: base, disambiguation_internal, disambiguation_user, hallucination_missing_tool, hallucination_missing_tool_parameter, hallucination_missing_tool_response
disambiguation_element_internal	string or null	What needs to be disambiguated internally (only set in disambiguation tasks)
disambiguation_element_user	string or null	What needs to be disambiguated with the user (only set in disambiguation tasks)
disambiguation_element_note	string or null	Note explaining the disambiguation (only set in disambiguation tasks)
removed_part	string (JSON) or null	Which tools/parameters were removed (only set in hallucination tasks)

Mock Data

The mock data simulates a realistic car environment database used by the tools during benchmark execution.

Config	Rows	Description
mock_locations	48	European cities with GPS coordinates
mock_pois	130,693	Points of interest (airports, bakeries, restaurants, etc.)
mock_weather	48	Weather data per location (8 time-slots/day)
mock_routes_location_location	6,768	Routes between locations (3 alternatives each)
mock_routes_location_poi	1,378	Routes from locations to POIs
mock_routes_poi_location	1,378	Routes from POIs to locations
mock_routes_index	1,763,870	Route lookup index
mock_routes_metadata	1,754,346	Metadata for POI-to-POI route generation
mock_calendars	100	Calendar entries with meetings
mock_contacts	100	Contact information

Usage

With the CAR-Bench benchmark

The CAR-Bench codebase loads tasks and mock data from this dataset automatically:

pip install -e .
python run.py --model gpt-4.1-mini --task-type base --task-split test --num-tasks 3

Standalone

from datasets import load_dataset

# Load tasks
tasks = load_dataset("johanneskirmayr/car-bench-dataset", "tasks_base")
print(tasks["test"][0])

# Load mock data
locations = load_dataset("johanneskirmayr/car-bench-dataset", "mock_locations", split="train")
contacts = load_dataset("johanneskirmayr/car-bench-dataset", "mock_contacts", split="train")

# Parse nested JSON fields
import json
task = tasks["test"][0]
context = json.loads(task["context_init_config"])
actions = json.loads(task["actions"])

Citation

If you use this dataset, please cite the CAR-Bench paper:

@misc{kirmayr2026carbenchevaluatingconsistencylimitawareness,
      title={CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty}, 
      author={Johannes Kirmayr and Lukas Stappen and Elisabeth André},
      year={2026},
      eprint={2601.22027},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.22027}, 
}