MiniWeb

MiniWeb is a benchmark for web agents that measures skills, not only tasks: 65 locally hosted websites, 37 interaction primitives (searching, filtering, filling forms, dragging, uploading, ...), and 376 human-written tasks whose every step is labeled with the primitive it exercises and checked by a verifier. MiniWebGen generates practice data for each primitive on the training sites, and MiniWebAgent is trained on it. The code that serves the sites, runs the evaluations and reproduces the paper is in the MiniWeb repository.

Tables

Config	Rows	One row per
tasks	376 (train 310, test 63, excluded 3)	human task: instruction, primitives, the task and verifier as evaluated, the annotator's originals, provenance, the demonstration's server and input logs
demonstrations	17.8K	recorded event of a human demonstration (action or observation), with the page's HTML, accessibility tree and screenshot where recorded; screenshot rows at the end hold captures no event points to
miniwebgen_trajectories	3,128	MiniWebGen trajectory of one primitive on a training site, with its backend and judge verdicts
miniwebgen_steps	17.1K	step of a MiniWebGen trajectory, with its screenshot and the targeted page element; phase is history for the steps of the primitive a chained trajectory starts with
training_steps	31K	step of MiniWebAgent's training data in the agent harness's format: prompt messages, target action, screenshot; is_target and validation mark the 27.6K steps it is fit on
task_centric_tasks	7,280	task proposed for the task-centric comparison, with its judged outcome
task_centric_steps	27.6K	step of the task-centric training data, in the same format as training_steps
single_primitive_test_tasks	115	generated single-primitive task of a held-out site (evaluation only)
sites	65	website: domain, train or held-out split, data source and its license
primitives	53	primitive (46 initial, 37 kept) or reasoning operation
guides	46	MiniWebGen's guide for a primitive

Nested fields (verifiers, logs, prompt messages, actions) are JSON strings. test is the 13 held-out sites: no training data comes from them.

from datasets import load_dataset

tasks = load_dataset("miniweb/MiniWeb", "tasks", split="test")
steps = load_dataset("miniweb/MiniWeb", "training_steps", split="train", streaming=True)

The websites

site_db/ holds everything the sites need to run offline: one Parquet file per table of the 65 sites (site_db/tables/<site>/), the schema (schema.json), the rows each full-text index holds (fts_rowids/), the sites' own images (static.parquet: generated product shots, post images, book covers, avatars), and a copy of the external media the pages show (media/: images that site data links to on other hosts, such as product and phone photos, wikiHow step images and images inside posts, plus the map's basemap tiles; media_manifest.parquet lists every linked URL and whether it could still be fetched). scripts/restore_site_db.py in the code repository downloads it, rebuilds the SQLite database the sites run on, with search behaving exactly as when the tasks were recorded, and unpacks the media, which the server then serves in place of the original links. Session state from our evaluations is not included.

Privacy

Annotators appear only as the pseudonyms A1 to A6. The sites' people are synthetic, apart from the public sources listed in sites (for example the Enron email corpus and the WebArena Reddit and GitLab environments).

License

• Data created for MiniWeb (tasks, verifiers, demonstrations, MiniWebGen and task-centric data, the synthetic data of 44 sites): CC BY 4.0.
• Data of the 21 sites seeded from public sources: the license of each source, listed in the sites table. Some are share-alike (CC BY-SA, ODbL) and some restrict use (wikiHow: CC BY-NC-SA; USA Real Estate: educational use), so the site database as a whole is for non-commercial research.
• Screenshots and page snapshots show site content and carry the license of that content.
• site_db/media/ archives third-party images that site data links to (for example product photos from WebShop's data and phone photos from GSMArena), as WebShop and WebArena use them, so the benchmark stays reproducible; their rights remain with their owners.