ziyjiang/WorldAuditBench-trajectories

WorldAuditBench Trajectories

Paper · Code · Project · Dataset

Exploration trajectories from GPT-6 Astra (VLM) on 213 WorldAuditBench tasks: 126 Unreal tasks and 87 Three.js tasks. Each task contains 40 actions, 41 screenshots, the agent's final report, and the judge result. In total: 8,520 actions and 8,733 screenshots.

Files

One archive per experiment. Extract gpt-6-astra-vlm.tar.gz to get:

gpt-6-astra-vlm/
├── A01/
│   ├── trajectory.json
│   └── screenshots/
│       ├── 000.jpg
│       ├── 001.jpg
│       └── ... 040.jpg
├── A02/
├── ...
└── JS_WT17/

000.jpg is the initial observation. 001.jpg through 040.jpg are the observations after each action. Image paths in the JSON are relative to the task folder.

trajectory.json

Field	Contents
task	Task ID, such as A01.
input	The experiment's instruction and scene description.
agent_final_report.final_response	The agent's complete final response.
agent_final_report.summary	The summary submitted for judging.
agent_final_report.bugs	Submitted findings, including descriptions, categories, status, and evidence references. An empty list means no findings were submitted.
judge_output	The original score and its explanation in English. 1 means the target anomaly was identified; 0 means it was not.
initial_observation	Path to the initial screenshot.
steps	Ordered actions and their resulting observation images.

For example, the first two steps of A01 are:

[
  {
    "step": 1,
    "action": {"kind": "turn", "params": {"degrees": 90.0}},
    "observation": "screenshots/001.jpg"
  },
  {
    "step": 2,
    "action": {"kind": "move", "params": {"distance_m": 4.0, "direction": "forward"}},
    "observation": "screenshots/002.jpg"
  }
]

Actions include move, turn, look, wait, and interact. Their parameters are stored in action.params.

Report references such as a23 correspond to screenshots/023.jpg. References such as a28#c1 denote crops used in the original report; separate crops and intermediate video frames are not included in this export. Judge references to numbered figures refer to the original judging evidence.

Screenshots and reports

Unreal screenshots were rendered again along the recorded actions using the released environments without the minimap mask. Three.js screenshots are the original observations. Actions, agent reports, and judge scores are preserved; judge explanations are translated into English. The historical instructions are retained, including references to the original HUD mask. Judging was not rerun on the new screenshots.

Download

hf download ziyjiang/WorldAuditBench-trajectories gpt-6-astra-vlm.tar.gz --repo-type dataset --local-dir .
tar -xzf gpt-6-astra-vlm.tar.gz

import json
from pathlib import Path

folder = Path("gpt-6-astra-vlm/A01")
trajectory = json.loads((folder / "trajectory.json").read_text())

print(trajectory["agent_final_report"]["final_response"])
print(trajectory["judge_output"])
for step in trajectory["steps"]:
    image_path = folder / step["observation"]
    print(step["step"], step["action"], image_path)