README.md

metadata

license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en
tags:
  - code
  - android
  - mobile
  - benchmark
  - swe-bench
  - kotlin
  - dart
  - flutter
  - react-native
  - issue-resolution
pretty_name: MobileDev-Bench
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.jsonl

MobileDev-Bench

MobileDev-Bench is a benchmark of 415 verified real-world issue-resolution tasks collected from 19 production mobile app repositories spanning Android Native (Java/Kotlin), React Native (TypeScript), and Flutter (Dart). It evaluates the ability of large language models to resolve authentic developer-reported issues in mobile applications through execution-based validation.

Paper: MobileDev-Bench: A Benchmark for Issue Resolution in Mobile Application Development

GitHub: mobiledev-bench (evaluation harness, scripts, Docker environments)

---

📊 Key Facts

Tasks	415 verified instances
Repositories	19 production mobile apps
Frameworks	Android Native · React Native · Flutter
Languages	Kotlin · Java · TypeScript · Dart
Difficulty tiers	Easy (139) · Medium (140) · Hard (136)
Avg. files changed	13.5
Avg. lines changed	344.2
Multi-artifact fixes	42.2% of instances
Human verified	✅ All instances manually verified

---

📖 Dataset Description

Motivation

Existing issue-resolution benchmarks (SWE-bench and its extensions) target general-purpose Python/JavaScript libraries. Mobile app development imposes distinct engineering constraints: platform-defined lifecycles, framework-specific build systems (Gradle, pub, npm), and correctness that depends on coordinated changes across heterogeneous artifacts — source files, manifests, resource definitions, and configuration metadata. MobileDev-Bench is the first benchmark to evaluate LLMs on these challenges at scale.

Construction

Each instance was derived from a merged GitHub pull request that resolves at least one linked issue. The pipeline:

1. Crawl — Pull requests and linked issues collected from 19 repositories (≥400 GitHub stars each).
2. Filter — Instances with no linked issue, no test changes, or non-compilable base commits are excluded.
3. Verify — Test execution under three configurations (base commit, test patch only, full fix) confirms that the fix produces a genuine Fail-to-Pass or None-to-Pass transition.
4. Annotate — Human annotators assign difficulty tiers (Easy / Medium / Hard) and label artifact types involved in each fix.

Splits

Split	Instances	Description
test	415	Verified benchmark instances with full evaluation metadata (fix_patch, test_patch, test_command, test results)

---

🗂️ Field Descriptions

Field	Type	Description
instance_id	string	Unique identifier: {org}__{repo}-{number}
org	string	GitHub organisation (e.g. thunderbird)
repo	string	GitHub repository name (e.g. thunderbird-android)
number	int	Pull request number
state	string	PR state at collection time (closed)
title	string	Pull request title
body	string	Pull request description
base	string	JSON object with base branch label, ref, and SHA
resolved_issues	string	JSON array of linked issue objects (number, title)
lang	string	Primary language: kotlin, java, dart, or typescript
problem_statement	string	Concatenated issue text — the task description shown to the model
hints	string	Concatenated issue comments providing additional context
pull_url	string	Full GitHub URL of the pull request
issue_urls	string	JSON array of linked issue URLs
fix_patch	string	Gold unified diff of the fix (applied to base commit)
test_patch	string	Unified diff adding or modifying tests that validate the fix
test_command	string	Shell command to execute the relevant test suite
fixed_tests	string	JSON list of test names that transition from Fail/None to Pass
f2p_tests	string	JSON list of Fail-to-Pass test names
p2p_tests	string	JSON list of Pass-to-Pass test names
s2p_tests	string	JSON list of Skip-to-Pass test names
n2p_tests	string	JSON list of None-to-Pass test names
run_result	string	Outcome of test execution on the base commit
test_patch_result	string	Outcome with test patch only (no fix)
fix_patch_result	string	Outcome with both test patch and fix applied

---

🏠 Source Repositories

Repository	Language	Link
zulip/zulip-flutter	Dart	GitHub
PalisadoesFoundation/talawa	Dart	GitHub
AntennaPod/AntennaPod	Java	GitHub
Futsch1/medTimer	Java	GitHub
thunderbird/thunderbird-android	Kotlin	GitHub
wordpress-mobile/wordpress-android	Kotlin	GitHub
element-hq/element-x-android	Kotlin	GitHub
streetcomplete/StreetComplete	Kotlin	GitHub
tuskyapp/Tusky	Kotlin	GitHub
commons-app/apps-android-commons	Kotlin	GitHub
openhab/openhab-android	Kotlin	GitHub
LemmyNet/jerboa	Kotlin	GitHub
PaulWoitaschek/Voice	Kotlin	GitHub
mjaakko/NeoStumbler	Kotlin	GitHub
JackEblan/Geto	Kotlin	GitHub
NMF-earth/nmf-app	TypeScript	GitHub
Expensify/App	TypeScript	GitHub
RocketChat/Rocket.Chat.ReactNative	TypeScript	GitHub
artsy/eigen	TypeScript	GitHub

---

🚀 Usage

from datasets import load_dataset

ds = load_dataset("MobileDev-Bench/mobiledev-bench", split="test")

Or clone the repository directly:

git lfs install
git clone https://huggingface.co/datasets/MobileDev-Bench/mobiledev-bench