VeriWorld

Executable test tasks for spatial reasoning in interactive 3D environments.

Code and setup · Windows and Linux binary · Croissant metadata

This dataset publishes task definitions and initialization parameters so users can run their own evaluations. It is not the authors' complete experiment-result dataset. The task source and runtime correspond to paper-artifacts-2026-11-03-rc1.

Configuration	Records	Unit
task_catalog	11	Task-family definition with objective, metric and code path
seed_fixtures	90	Reference generator configuration and initialization parameters
runtime_recipes	22	Representative CLI installation-check command

The configurations have different units; their row counts must not be summed to count scenes or model runs. test is the Hub split name for these reference records, not a newly defined training/test partition.

Load task definitions and seeds

import json
from datasets import load_dataset

revision = "c2e91ddba815b394db165518311d8d092ec2463f"
tasks = load_dataset("yanyanzheng96/VeriWorld", "task_catalog", split="test", revision=revision)
fixtures = load_dataset("yanyanzheng96/VeriWorld", "seed_fixtures", split="test", revision=revision)
parameters = json.loads(fixtures[0]["parameters_json"])

The pinned data revision preserves the exact task and parameter CSV bytes. The current croissant.json adds task-only scope and source/generation metadata. No custom remote code is required to read the CSVs. Reading metadata does not execute a task or produce a model score.

Run a task

Download and verify the binary and source using the GitHub release instructions. Follow the source README to install dependencies and configure your own model API. From the extracted source/bundle directory:

python run_task.py --check
python run_task.py --task PlaceOnBottom --seed 0 --model example-model --smoke
python run_task.py --task DropToTarget --seed 0 --condition V --model example-model --smoke

Replace example-model with your configured route. --smoke runs a one-round installation check; it is not a full benchmark evaluation. The runtime_recipes configuration supplies commands for all eleven task families. Exact action and verifier definitions are in the pinned task implementation.

Scope, provenance and responsible use

Offline fixtures cover eight families. LightWar, SceneGrounding and ObjectGrounding use live runtime/scene configuration; their task definitions and CLI recipes are included. Legacy TunnelNav3D uses construction seed 99 for rendered tunnel waypoints, so changing the CLI seed alone does not give an independent tunnel geometry. See TASK_SEED_NOTES.txt.

Task parameters can contain evaluation-side reference quantities. The harness defines which fields are provided in each input condition; do not append hidden references to image-based input. For PlaceOnBottom the V input also includes anonymous object identifiers with positions and sizes; the pure-visual (V-pure) condition supplies the image and identifiers only. Repeated model calls on one configuration are not independent seeds. API failures remain distinct from model FAIL outcomes. No API credentials are included.

DATASET_METADATA.md documents record units, source and generation activities, runtime dependencies, limitations and metadata revision. The canonical Croissant describes the three task configurations with immutable data URLs and SHA256 digests, plus required Responsible AI and provenance fields. Use that file when supplying author-maintained metadata; an automatically generated Hub Croissant may omit the added RAI/provenance explanations.

Historical score CSVs and SUPPORTING_DATA.txt remain available as separate paper-support files. They are excluded from the task dataset's Croissant and viewer configurations. They are neither a complete experiment archive nor a claim that later repair/expansion results have been published.

The benchmark uses synthetic scenes rather than human-subject or scraped personal data. Its performance does not establish safety or deployment readiness for robots or embodied systems. Live checks cover representative configurations, not every task × seed × model combination.

Licensing and citation

See LICENSE.md for the existing scopes of task/project material, historical numeric tables, and external runtime/assets. This metadata revision does not change those terms. Use CITATION.cff to cite the paper.