README.md

metadata

license: mit
language:
  - en
pretty_name: MyPCBench Tasks
size_categories:
  - n<1K
tags:
  - arxiv:2606.16748
  - computer-use
  - agents
  - benchmark
  - llm-as-judge
  - personal-assistant
  - desktop
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/tasks.jsonl

MyPCBench — Task Set

A benchmark for personally intelligent computer-use agents.

This dataset is the task set and grading rubrics for MyPCBench. It accompanies the VM image ljang0/mypcbench-qemu-baseline and the paper.

MyPCBench is a reproducible Linux-desktop benchmark seeded end-to-end from one canonical persona (Michael Scott, The Office). The image hosts 17 pre-logged-in web apps mirroring real consumer products plus the full LibreOffice suite; the persona's records (bank transactions, emails, calendar events, chat messages, retail/grocery/food orders, web visits, rides) are cross-linked, so one trip leaves correlated records across every app that would plausibly book it. Agents are evaluated on tasks that require knowing who the user is.

Loading

from datasets import load_dataset

ds = load_dataset("ljang0/MyPCBench-tasks")  # DatasetDict({test: 184 rows})
task = ds["test"][0]
print(task["instruction"])
print(task["grading"]["rubrics"])

Contents

• 184 tasks, all rubric-graded (LLM-as-judge).
• data/tasks.jsonl — the dataset split the viewer and load_dataset use. One row per task, with grading rubrics inlined and the paper's behavioural-type label merged in.
• raw/ — the verbatim source files from the GitHub repo, for fidelity:
  • all_tasks_with_grading.json — flat eval input (grading inlined).
  • all_tasks.json — convenience flat list without grading (do not grade with this).
  • task_types.json — per-task behavioural-type mapping (paper taxonomy).
  • variables.json — ground-truth values reference (in-VM paths, balances, …).
  • schema.json — task JSON schema.

Fields

Each row has:

field	description
id	unique task id, e.g. retrieval-f016
app	primary app the task targets
apps_involved	apps the task is expected to touch (analysis only, not enforced)
category	internal task category (retrieval, aggregation, contradiction, situated_action, preference_inference, counterfactual, long_horizon, cua_only, hard_app)
behavioral_type	paper's six-type taxonomy: bounded_action (64), multi_step_orchestration (48), cross_source_reconciliation (25), aggregation_reporting (23), personal_lookup (13), pattern_inference (11)
difficulty	easy / medium / hard
instruction	the natural-language task instruction given to the agent
num_rubrics	number of grading rubrics (3–13 per task)
grading	{ "type": "llm_judge", "rubrics": [ { "criterion", "type", "weight" } ] } — rubric weights sum to 1.0

Grading

Grading is rubric-only (LLM-as-judge). The runner records episode completion; an offline judge scores each rubric over the full trajectory. See the GitHub repo for the runner and judge.

Citation

@misc{jang2026mypcbench,
  title  = {MyPCBench: A Benchmark for Personally Intelligent Computer-Use Agents},
  author = {Jang, Lawrence},
  year   = {2026},
  url    = {https://huggingface.co/papers/2606.16748}
}

License: MIT.