markov-ai/autocad-bench-tasks

AutoCAD-Bench Tasks

This dataset contains the 50 AutoCAD tasks used for model evaluation in AutoCAD-Bench.

Contents

• 50 total evaluation tasks
• 21 dimensioned 2D drafting tasks
• 29 3D modeling and presentation tasks
• millimetres as the task unit
• one reference PNG and one scorer-side gold DWG per task
• the exact 2D and 3D task prompts used by the benchmark

Each task is stored under data/task-NNN/:

data/task-001/
├── reference.png
└── gold.dwg

tasks.jsonl records the sequential task ID, drawing type, difficulty level, units, task prompt, artifact paths, file sizes, and SHA-256 checksums.

Evaluation boundary

The model should receive only the task prompt and reference.png. The gold DWG is an evaluator-side artifact and must never be mounted in the model-controlled desktop, included in the model prompt, or otherwise made available to the agent during a rollout.

For 2D tasks, the expected submission is a complete dimensioned drawing in Model Space. For 3D tasks, the expected submission contains genuine 3D geometry in Model Space and the complete reference-sheet presentation on Layout1.

Loading the manifest

import json
from pathlib import Path

tasks = [
    json.loads(line)
    for line in Path("tasks.jsonl").read_text().splitlines()
    if line.strip()
]

The DWG files use the AutoCAD 2018 AC1032 file format.