YuvrajSingh9886/androidlife-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, difficulty, app(s), ASK USER flags, and prompt text.

Config	File	What you see
tasks (default)	benchmark/tasks_preview.jsonl	Compact preview columns for browsing
tasks_full	benchmark/AndroidLife_public_v2.jsonl	Full per-task records (placeholders, templates, …)

benchmark/ — public task sample

File	Content
AndroidLife_public_v2.json / .jsonl	60 public tasks (20/day × 3 days; Easy/Medium/Hard).
tasks_preview.jsonl	Compact rows for the HF Dataset Viewer.
public.md	Human-readable source prompts.
ask_user_facts_public.json	Withheld facts for ASK USER SINGLE tasks (7).
multiturn_kb_public.json	KB profiles for ASK USER - MULTI tasks (4).
hallucination_controls.json	Hallucination-control task specs used by the public sample.
public_vars.example.env	Placeholder template (safe to copy).
public_vars.local.env	Resolved placeholder values used on the published test device.

Run artifacts

• runs/<run_id>/ — raw execution artifacts (agent logs, LLM proxy metrics, telemetry, per-step GUI evidence, trajectories)
• reports/public/ — per-run narrative reports (manual audit, verdicts, on-device verification)
• reports/metrics/public/ — per-run official metrics + manual-audit notes (manual audit is ground truth)
• reports/metrics/hallucination/public/ — hallucination-control judge outputs
• reports/turn-based/public/ — per-turn ASK USER audits

Per-task files inside runs/<run_id>/

File	Content
agent.log.txt	Full agent step log (MobileRun FastAgent).
output.json / output.txt	Graded outcome (success + reason) and final answer.
meta.json	Run metadata (task_id, model, command, exit code).
preflight.json / postflight.json	Device pre/post checks.
llm_proxy_metrics.jsonl	Per-LLM-call tokens + cost.
ask_user_metrics.jsonl	ask_user tool calls (ASK USER tasks).
samples.ndjson	Battery/thermal sampling during the task.
trajectories/<ts>/	Per-step GUI evidence (macro.json, trajectory.json, ui_states/, screenshots/).

Fabricated test data

All on-device seed data is fabricated for the benchmark (fictional persona contacts, notes, docs, etc.). Numbers/emails visible in trajectories are synthetic seeds, not real personal data. See the fabrication/ disclosure in androidlife-530.

License & citation

• Dataset / run artifacts / tasks: CC BY 4.0 — free to use and adapt (including commercially) with attribution to Yuvraj Singh. Do not remove credit.
• Harness code: Apache-2.0 in the AndroidLife repo.

If you use AndroidLife — the benchmark, leaderboard, tasks, trajectories, or results — please credit this work and cite it as:

@misc{singh2026androidlife,
      title={AndroidLife: Real-Phone Android Agent Benchmark for Open-Weight Models and On-Device SLMs},
      author={Yuvraj Singh},
      year={2026},
      howpublished={\url{https://github.com/YuvrajSingh-mist/AndroidLife}},
}

---

AndroidLife — a real-phone Android agent benchmark. 530 tasks, graded on the on-device end state: not a simulator, not a mock API.

Leaderboard + step-by-step replays · All models & datasets · Code

Fuel the next run — a full 60-task suite is ~6.5 h of phone time plus real API spend: