CUA-Bench: 75 original CUA-Gym tasks

A selection of 75 unchanged original tasks from xlangai/CUA-Gym, grouped into three evaluation splits of 25 tasks each. Every task includes its original instruction, setup script and matching reward.py. No synthetic task variants or replacement grading rules were created.

This repository also includes the source code for the 13 required mock websites, a Docker environment, and a controller helper for running original setup and reward code. The browser agent can use DOM observations; grading uses the original task's native state-based verifier. A model and its agent loop are supplied by your evaluation harness.

Current execution status

This release is not a fully validated 75-task benchmark. All 75 original setups and verifiers executed, but only 54/75 pages passed a DOM-content check. 21/75 pages failed to render usable content in the unchanged pinned upstream apps: 16 Uber Eats and 5 Monday tasks. Preserve these as environment failures, not model failures. See KNOWN_ISSUES.md for exact task IDs and evidence.

One original Pinterest task was completed using DOM actions and received 1.0 from its original verifier, after a no-action score of 0.0. No 9B-model rollouts were run. Downloadable task/verifier completeness and successful task solvability are separate checks.

Splits

Split	Tasks	Included workflows
navigation_view_settings	25	Open views, browse stores/stories, switch accounts or document modes, change display/calendar settings
search_workflows	25	Search, sort and filter; some tasks also open a result, add cart items or favorite results
record_form_create	25	Create contacts, events, tasks, folders, groups, notes, posts, pins, addresses and other bounded records
Total	75	75 distinct original source task IDs

Category scope: navigation includes settings; search includes multi-step discovery and result actions. This is not a collection of 25 pure navigation and 25 single-query tasks. All original requirements remain intact. Search has 15 Uber Eats, 4 Instacart, 3 Pinterest and 3 Monday tasks; site diversity is uneven.

The selection favors bounded, single-app tasks. Difficulty has not been calibrated for a 9B model, and original source difficulty labels are retained in task.json. Original dates, named targets, field values and image URLs are preserved.

Load the splits

from datasets import load_dataset

ds = load_dataset("ishagarg1103/cua-bench")
assert {name: len(split) for name, split in ds.items()} == {
    "navigation_view_settings": 25,
    "search_workflows": 25,
    "record_form_create": 25,
}

Loading the rows alone does not download the task or website archives. Download the complete repository to execute tasks:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ishagarg1103/cua-bench",
    repo_type="dataset",
    local_dir="cua-bench",
)

Included execution assets

Path	Contents
data/*.jsonl	Three task splits, instructions and exact artifact references
artifacts/original_tasks.tar.zst	Original <task-id>/task.json, initial_setup.py, reward.py for all 75 tasks
artifacts/cua_gym_hub_apps.tar.gz	Pinned source for all 13 mock apps, shared server code, original npm lockfiles and upstream license notices
scripts/runtime.py	List tasks, verify/extract originals, materialize runtime URLs, run setup and invoke the original verifier
environment/Dockerfile	Python, Node, Chromium, Xvfb and app installation recipe
environment/requirements.txt	Python execution dependencies
environment/lock.json	App revision, file hashes, app mapping and per-task deployment origins
upstream/	Original upstream URL-materialization script, full URL manifest and source dataset README
provenance.json	Original task-source revision, IDs and per-file hashes
validation.json, validation/	Packaging checks, per-task startup results, raw log archive and one successful DOM reference
KNOWN_ISSUES.md	Exact task IDs affected by upstream rendering failures and nonzero no-action rewards
EXECUTION.md	Docker/Daytona execution and DOM harness integration

Task row fields

Field	Meaning
id	Unchanged upstream task ID
instruction	Unchanged upstream agent instruction
app_type	Original mock application name
archive_path	Task bundle archive, relative to this repository
archive_member	Task folder inside that archive
task_json_member	Original task definition inside the archive
reward_member	Matching original verifier inside the archive
setup_file_members	Matching original setup artifacts inside the archive
initial_url_template	Original browser start URL with {sid} to be filled after setup; requires deployment URL substitution
url_variable	Upstream environment variable for the app's deployment origin

Verification contract

The original task definition contains:

{"evaluator": {"type": "python", "url": "./reward.py"}}

The setup creates a fresh application state and writes its session ID to /tmp/task_web_sid. The browser agent acts in that same session. The controller runs the corresponding reward.py, which reads app state and prints REWARD: <score>. Keep the native fractional score; this release does not replace it with a custom binary grader.

Original files remain byte-identical in the distributed archive. Runtime copies substitute deployment endpoints only. Two upstream tasks contain literal private-IP endpoints instead of placeholders; their exact origin substitutions are listed in environment/lock.json and handled by the runner. Target content URLs and scoring logic are unchanged.

Validation scope

Source hashes, unique IDs, original instructions, setup/verifier pairing, complete task bundles, Python syntax and URL materialization were checked for all 75 tasks. Runtime coverage: 75 original setup runs, 75 no-action native reward calls, 54 DOM-content checks passed, 21 rendering failures, and one successful DOM reference completion. See the machine-readable receipts for details. A startup/no-action smoke test does not establish that a task is solvable or that its reward is correct after successful completion. Full successful DOM completion of every task and 9B-model evaluation remain unvalidated.

The helper and deployment metadata are new integration code. The 75 tasks, setup logic and reward logic are from the original dataset. Keep setup/verifier files and state APIs under the trusted controller; expose the task instruction and browser observations/actions to the evaluated model.

Sources and licenses

Original task content: xlangai/CUA-Gym, revision 3c021d06e2b01bbb6ad28cfddec41d856e4a78c5, by XLang NLP Lab / xlang-ai, licensed CC BY 4.0. Original content is unchanged; selection, workflow labels and packaging metadata were added.

Mock website code: xlang-ai/CUA-Gym-Hub, revision 53205689c3d88078c1375f76466d5bd799478828. Its Apache-2.0 license, NOTICE and TRADEMARKS.md are included in the website archive. This app revision is a pinned deployment choice, not a claim that upstream published an exact task-to-Hub revision mapping. The sites are local research mocks, not the real commercial services.