Hugging Face's logo

Datasets:
TAG-Nav
/
DestBench

Tasks:
Image-to-Text
Modalities:
Image
Formats:
imagefolder
Languages:
English
Size:
< 1K
Tags:
gui-agents
mobile
android
navigation
benchmark
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1
DestBench / README.md
TAG-Nav's picture
TAG-Nav
Table cleanup
59806d7 verified
preview code
|
Raw History Blame Contribute Delete
2.42 kB
metadata 
license: cc-by-nc-4.0
pretty_name: DestBench
task_categories:
  - image-to-text
language:
  - en
tags:
  - gui-agents
  - mobile
  - android
  - navigation
  - benchmark
size_categories:
  - n<1K

DestBench (core-6 release)

DestBench is a destination-reaching benchmark for mobile GUI agents: given a natural-language instruction, the agent must reach a specific app page. It accompanies the AACL-IJCNLP 2026 Findings paper Not All Graphs Are Maps: A Task-Agnostic Exploration Framework for Mobile GUI Navigation (TAG-Nav).

Project page: https://tag-nav-authors.github.io/tag_nav/

Contents

This release contains the 84 tasks of the six core apps used in the paper's main evaluation:

App folder App App version Tasks
google_maps_26.06.01.863982022 Google Maps 26.06.01 24
amazon_32.3.0.100 Amazon Shopping 32.3.0.100 17
youtube_21.05.264 YouTube 21.05.264 15
calendar_2026.05.0-864040481-release Google Calendar 2026.05.0 10
settings_16 Android Settings Android 16 9
clock_8.5 Google Clock 8.5 9

Each task folder <app>/<task>/ holds:

  • frame_XXXXXX_*.png: the reference trajectory, one screenshot per step. The last frame is the destination (target) page.
  • frame_XXXXXX_*.xml: the UI hierarchy dump for that step, where available.
  • meta.json: prompt (the evaluated instruction), prompts (paraphrases), and actions (the reference tap/scroll sequence with screen coordinates).
  • prompts.json: the instruction paraphrases.

Privacy

The trajectories were recorded on test devices. Every screenshot and UI dump was screened for personal data, by OCR and by eye. Account names, email addresses, profile pictures, device identifiers, addresses and user-generated content were removed with solid fills in the images and placeholders in the XML (user@example.com, [place], [review text removed]). Filled regions are intentional and do not affect the destination page. If you find anything that identifies a person, please open a discussion on this repository and we will remove it.

Intended use

For research on GUI agents, evaluated on emulators, test devices and test accounts. Do not use it to automate actions on accounts you are not authorized to use, or to collect data about app users. Screenshots of third-party apps remain subject to the respective app vendors' rights.