datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
AndroidtvasstestAndroid-in-the-Wild
Android in the Wild (AITW)
This is a mirror of Google's Android in the Wild (AITW) dataset, re-hosted on Hugging Face for easier community access.
Original Source
Paper: Android in the Wild: A Large-Scale Dataset for Android Device Control
Original Repository: google-research/google-research/tree/master/android_in_the_wild
Dataset Description
Android in the Wild (AITW) is a large-scale dataset for Android device control. It contains human demonstrations of… See the full description on the dataset page: https://huggingface.co/datasets/leosltl/Android-in-the-Wild.androidlife-530
AndroidLife-530 — Android agent benchmark (real phone, real LLM)
AndroidLife runs Android agent tasks against a real phone (via ADB/MobileRun)
and a real LLM, and grades the agent on reaching a verifiable device end-state.
This repo ships the 530-task corpus plus everything needed to reproduce runs.
Benchmark, or template — your call. The 530 tasks are an extended version
of the benchmark, usable as a larger evaluation set for further benchmarking of
models beyond the 60-task… See the full description on the dataset page: https://huggingface.co/datasets/YuvrajSingh9886/androidlife-530.android-controlAndroidAVDAndroidCodeandroidlife-public
AndroidLife — Public 60-task preview + run trajectories
This dataset publishes the public 60-task AndroidLife sample (task definitions +
sidecars) together with the full per-run execution artifacts from real-phone
benchmarks (OnePlus CPH2423 via ADB/MobileRun + real LLMs).
The full 530-task corpus lives in
YuvrajSingh9886/androidlife-530.
Dataset preview (tasks)
Open the Dataset Viewer above (config tasks) for a table of all 60 public
tasks: task_id, day… See the full description on the dataset page: https://huggingface.co/datasets/YuvrajSingh9886/androidlife-public.Android-Malware-Datasetandroidlife-trajectories
AndroidLife300 — trajectory replays
Public trajectory assets for the AndroidLife Android-agent benchmark, served to the published site via resolve/main URLs.
index.json — per-task availability manifest (model, success, steps, gif/data URLs)
data/trajectories/<set>/<day>/<task>.json — condensed step streams (thoughts, tool calls, screenshots)
trajectories/<set>/<day>/<task>/trajectory.gif + screenshots/ — screen replay + per-step frames
Regenerate/republish with uv run python… See the full description on the dataset page: https://huggingface.co/datasets/YuvrajSingh9886/androidlife-trajectories.AndroidFlux_Failure_Recovery_Eval
AndroidFlux Failure Recovery Eval
The current replay-verified benchmark release is v2, with 65 tasks and one
Docker-replay-verified internal-error trajectory per task. Replay bundles and
their model-ready history metadata live under the explicit v2/ prefix.
The files under v1/, v1.1/, and the root data/ prefix are retained for
reproducibility. Do not combine globs from different releases.
v2 contents
v2/data/replay/*.parquet: 65 replay bundles, including… See the full description on the dataset page: https://huggingface.co/datasets/Gyubeum/AndroidFlux_Failure_Recovery_Eval.AndroidFlux_results
🚨 STATUS / INCIDENTS
08-30 11:00 KST — RESOLVED — Root cause of the ~2h full stall (found after Bash access itself broke): the v2_65_t_minus_n results directory (episode screenshots) was writing to the host's small 49GB root disk instead of the 5TB /data2 volume, and filled it to 100%, which cascaded into everything (my own tool's tmp dir, new docker containers, vLLM server startups, the automated queue's next stage) failing with no-space errors. Moved the 9.8GB results dir to… See the full description on the dataset page: https://huggingface.co/datasets/Gyubeum/AndroidFlux_results.android_controlAndroidControl-Star
AndroidControl*
A curated step-level evaluation subset extracted from AndroidControl, used for static mobile GUI understanding evaluation in UI-MOPD (Multi-platform On-Policy Distillation for Continual GUI Agent Learning).
Overview
AndroidControl* contains 4,260 step-level records from 781 Android trajectories. Each record includes the trajectory identifier, step index, high-level task goal, per-step instruction, normalized action, screenshot path, screenshot… See the full description on the dataset page: https://huggingface.co/datasets/UI-MOPD/AndroidControl-Star.Android-Control
Android Control (Community Mirror)
This is a community mirror of the official Android Control dataset by Google Research, hosted on Hugging Face for easier access.
The original dataset is hosted on Google Cloud Storage at gs://gresearch/android_control/. This copy is provided to make it more accessible to the research community via Hugging Face.
Dataset Description
Android Control is a dataset for training and evaluating Android device control agents. It contains 15… See the full description on the dataset page: https://huggingface.co/datasets/leosltl/Android-Control.AndroidFlux_Evaluation_Outputandroid_controlAndroidControlParsedWithImages-20kcqadupstack-android
CQADupstackAndroidRetrieval
An MTEB dataset
Massive Text Embedding Benchmark
CQADupStack: A Benchmark Data Set for Community Question-Answering Research
Task category
t2t
Domains
Programming, Web, Written, Non-fiction
Referencehttp://nlp.cis.unimelb.edu.au/resources/cqadupstack/
How to evaluate on this task
You can evaluate an embedding model on this dataset using the following code:
import mteb
task = mteb.get_tasks(["CQADupstackAndroidRetrieval"])… See the full description on the dataset page: https://huggingface.co/datasets/mteb/cqadupstack-android.AndroidControl_3000_samples_trajectoryAndroidLens
AndroidLens: Long-latency Evaluation with Nested Sub-targets for Android GUI Agents
AndroidLens is a challenging benchmark for mobile GUI agents, featuring 571 real-world, long-horizon tasks in both Chinese and English, with an average of 26.1 steps per task. It supports evaluation of critical capabilities:
Long-horizon planning under multi-constraint & multi-goal scenarios
298 cross-app tasks and 273 single-app tasks, covering 74 real-world applications (e.g., WeChat, Google… See the full description on the dataset page: https://huggingface.co/datasets/yuecao0119/AndroidLens.android-firmware-research
Android Firmware & OTA Research Dataset
This dataset contains raw firmware dumps, OTA updates, and system images (like boot.img, vendor_boot.img, and full system zips).
It is intended for hardware security research, kernel patching, and Android vulnerability analysis.
wakeforge-hey-android-piper-tts
hey_android — Wake Word Synthetic Speech Dataset
Synthetic, augmented audio for training a small wake-word / keyword-spotting
model. Generated with piper_tts and local audio augmentation.
Classes
Label
Samples
background_noise
200
hey_android
378
unknown
1071
hey_android — the target wake phrase and close variants.
unknown — near-miss and unrelated short phrases.
background_noise — synthetic background noise.
Audio Specification… See the full description on the dataset page: https://huggingface.co/datasets/eoinedge/wakeforge-hey-android-piper-tts.AndroidControlAndroidControl_Curated
AndroidControl-Curated: Revealing the True Potential of GUI Agents through Benchmark Purification
This is the official repository for the paper AndroidControl-Curated.
🌟 Overview
On-device virtual assistants like Siri and Google Assistant are increasingly pivotal, yet their capabilities are hamstrung by a reliance on rigid, developer-dependent APIs. GUI agents offer a powerful, API-independent alternative, but their adoption is hindered by the… See the full description on the dataset page: https://huggingface.co/datasets/batwBMW/AndroidControl_Curated.Android-World-Eval
Android-World-Eval: RealGUI Evaluation Results
This dataset contains the evaluation results of Xiaomi-GUI-0 method on the Android World benchmark.
Overview
Method: Xiaomi-GUI-0
Benchmark: Android World (116 tasks)
Rollouts: 4 independent evaluation runs per task (464 total samples)
Results
Rollout
Samples
Success
Accuracy
0
116
94
81.03%
1
116
91
78.45%
2
116
91
78.45%
3
116
90
77.59%
Overall
464
366
78.88%
Data… See the full description on the dataset page: https://huggingface.co/datasets/SeerRay-Lab/Android-World-Eval.My-Saved-Android-DownloadsAndroidFlux_RL_Train
AndroidFlux RL — policy prompts
Slice
Prompts
Source
t_minus_n
2,459
AndroidFlux source trajectories
successful
2,623
AndroidFlux source trajectories
t_minus_1
2,459
AndroidFlux source trajectories
t
2,459
AndroidFlux source trajectories
Subtotal
10,000
ui_genie
10,000
UI-Genie reward-model prompts (5,000-prompt core marked by in_reduced)
The four AndroidFlux slices are drawn from replayed source trajectories.
The successful slice contains 53… See the full description on the dataset page: https://huggingface.co/datasets/Gyubeum/AndroidFlux_RL_Train.AndroidDaily
AndroidDaily Dataset
This repository hosts the AndroidDaily dataset, a benchmark grounded in real-world mobile usage patterns, introduced in the paper Step-GUI Technical Report.
The AndroidDaily benchmark comprises 3146 static actions and 235 end-to-end tasks across high-frequency daily scenarios. It is specifically designed to assess whether GUI agents can handle authentic everyday usage, providing a robust evaluation for GUI automation capabilities.
Project Page:… See the full description on the dataset page: https://huggingface.co/datasets/stepfun-ai/AndroidDaily.marvin_paranoid_androidThis dataset contains 15 images of Marvin, the paranoid android from the movie "The Hitchhiker's Guide to the Galaxy" (2005) scraped from the Internet and 205 images of general robots, created with Stable Diffusion from the prompt "a photo of a robot".
android_control_test