MazenBaShammakh/gui-failure-suite

GUI Failure Suite

Consolidated benchmark suite of 2,821 GUI agent tasks spanning mobile, web, and desktop platforms. Duplicate tasks across benchmarks have been removed. Includes 128 hand-authored custom tasks (GUIFailureSuite benchmark) on top of the 2,693 source-benchmark tasks.

---

Getting started

The suite is data only — there's no code to run, just files to load. It ships as three parallel, always-in-sync serializations of the same 2,821 records (csv/, json/, jsonl/), split into the three chunks below; concatenate them for the full set.

Load the full suite with pandas:

import pandas as pd
import glob

df = pd.concat(pd.read_csv(f) for f in sorted(glob.glob("csv/gui_suite_*.csv")))
print(len(df))  # 2821

Or line-by-line from the JSONL chunks:

import json
from pathlib import Path

records = []
for chunk in sorted(Path("jsonl").glob("gui_suite_*.jsonl")):
    with open(chunk, encoding="utf-8") as f:
        records.extend(json.loads(line) for line in f)

---

Files

The suite is distributed across 3 chunks of up to 1,000 records each, available in three formats:

File	Records	ID range
gui_suite_1	1,000	aitw-mobile-0001 → mind2web-web-1000
gui_suite_2	1,000	mind2web-web-1001 → mind2web-web-2000
gui_suite_3	821	mind2web-web-2001 → gui-failure-suite-mobile-2821

Each chunk exists in three formats under the corresponding subdirectory: json/, jsonl/, and csv/.

---

Record schema

Field	Type	Description
id	string	Unique task ID — {benchmark}-{platform}-{XXXX} (4-digit counter)
benchmark_id	string	Sub-split identifier within the source benchmark
benchmark	string	Source benchmark name
split	string	Dataset split (always test)
app	string | null	Target software application or website; null for AITW (not annotated)
platform_type	string	Platform: mobile, web, desktop, or desktop_windows
task	string	Natural-language task instruction

---

The app field

app is a machine-launchable identifier for the task's target, in one of four formats depending on platform_type. Multi-app entries are comma-separated with the primary app first (e.g. "com.android.chrome, com.google.android.googlequicksearchbox"); a consuming harness should open the first one, falling back to the next if it fails to launch. app is null only for AITW mobile tasks, where the source benchmark doesn't annotate a target app.

platform_type	app format	Example
web	HTTPS URL	https://www.seatgeek.com
mobile	Android package name	com.google.android.apps.tasks
desktop	Linux shell command	libreoffice --calc
desktop_windows	Windows executable path/command	C:\Program Files\Microsoft Office\root\Office16\EXCEL.EXE

---

Composition

By benchmark

Benchmark	Tasks	Platform	app field
AITW	497	mobile	null (not available in source)
LlamaTouch	477	mobile	app name string
Mind2Web	1,341	web	website name string
OSWorld	378	desktop / desktop_windows	app name(s), comma-separated
GUIFailureSuite	128	mobile / web / desktop_windows	app/package name, default per platform
Total	2,821

By platform

Platform	Tasks
mobile	1,064
web	1,375
desktop	358
desktop_windows	24

---

Source datasets

Benchmark	Full name	Platform	Reference
AITW	Android in the Wild	mobile	He et al., 2023 — arxiv.org/abs/2307.10088
LlamaTouch	LlamaTouch	mobile	Zhang et al., 2024 — arxiv.org/abs/2404.16054
Mind2Web	Mind2Web	web	Deng et al., 2023 — arxiv.org/abs/2306.06070
OSWorld	OSWorld	desktop (Linux & Windows)	Xie et al., 2024 — arxiv.org/abs/2404.07972
GUIFailureSuite	GUI Failure Suite (custom)	mobile / web / desktop_windows	Hand-authored tasks targeting known GUI failure modes

---

License & attribution

This suite is a derivative aggregation of four external benchmarks — AITW, LlamaTouch, Mind2Web, and OSWorld — deduplicated and normalized to a common schema, plus 128 hand-authored tasks. It is provided as-is for research use; consult each source benchmark's own license and terms (linked above) before redistributing tasks drawn from it. If you use this suite, please also credit the original benchmark authors for any tasks sourced from their work.