OpenJev-Data

Jun Huang*, Xin Ren* · University of Electronic Science and Technology of China · *Equal contribution

OpenJev (project wev) is an open-source, local alternative to Jev's System One decision API: same request shape, your own GPU, no API key.

Browser-step data and a benchmark for decision models: the small, fast models that pick a web agent's next operation and target element. Every row is one browser step, written as a POST /v1/systemone request (the page state plus typed questions) with its labelled answers. The requests use exactly the format the open browser agent jev-ultrafast sends to its System One, so a model trained or scored here can be served behind that agent unchanged. The data trains the wev models (wev-1.7b, wev-4b, wev-8b), and the bench config is a frozen test set, wev-bench, that scores any System One endpoint.

Load it

Every config is parquet and loads with one call:

from datasets import load_dataset

m2w      = load_dataset("alanhuangya/OpenJev-Data", "mind2web")          # human steps, split by website
nnetnav  = load_dataset("alanhuangya/OpenJev-Data", "nnetnav_audited")   # live-web steps with audited DONE labels
webchain = load_dataset("alanhuangya/OpenJev-Data", "webchain")          # human steps on 200+ live sites
qwen     = load_dataset("alanhuangya/OpenJev-Data", "teacher_qwen3max")  # qwen3-max acting as System One
glm      = load_dataset("alanhuangya/OpenJev-Data", "teacher_glm")       # GLM-5.3-Flash acting as System One
onpolicy = load_dataset("alanhuangya/OpenJev-Data", "onpolicy_glm")      # student-visited states, GLM labels
tasks    = load_dataset("alanhuangya/OpenJev-Data", "live_tasks")        # goals + start URLs; test = 153 held-out tasks
bench    = load_dataset("alanhuangya/OpenJev-Data", "bench", split="test")  # wev-bench, frozen

Configs

Config	Rows (train / validation / test)	Source	License
mind2web	5,861 / 586 / 875	Mind2Web, converted	CC BY 4.0
nnetnav_audited	11,408 / 1,140 / 1,150	NNetNav-live, converted, stopping labels audited	Apache-2.0
webchain	17,876 / 1,069 / 1,589	WebChain v2, converted from its HTML snapshots	CC BY 4.0
teacher_qwen3max	4,781 / 385 / 57	qwen3-max as the System One of jev-ultrafast on live sites (two collections)	see Terms
teacher_glm	5,969 / 370 / 147	GLM-5.3-Flash in the same role	see Terms
onpolicy_glm	5,782 / 275 / 119	DAgger-style: a wev student or the GLM teacher acts, the GLM teacher labels	see Terms
live_tasks	1,402 / – / 153	goals and start URLs for live runs; test is the held-out end-to-end suite	see Terms
bench	– / – / 2,346	wev-bench: frozen test steps from the configs above	per row's source

In v1 the two qwen3-max collections were separate configs (teacher_first, teacher_second); they are now one config with a collection column (teacher-v1, teacher-v2). v1 shipped JSONL; v2 ships parquet.

The general typed-decision corpora used in training (Kev decision-v7, typed-decisions, tasksource-jev, jev-distill-corpus-v3, typed-decisions-synth) are not redistributed here; the repository's builders download and convert them from their sources.

Columns

Step configs (all but live_tasks) share one schema:

Column	Meaning
request	the full POST /v1/systemone request as a JSON string; send it as is to replay the step
labels	JSON string: the gold option key per labelled question, e.g. {"operation": "CLICK", "click_target": "7"}
meta	JSON string: the source's own metadata (for teacher rows: episode, task, teacher, its one-sentence reason)
id	stable row id (hash of request and labels)
url, title, goal	page URL and title, task goal
n_elements, n_recent_actions	candidate elements offered (8–40 for converted data), actions in the history
operation	gold operation: CLICK, TYPE_TEXT, SELECT, SCROLL_DOWN, SCROLL_UP, WAIT, DONE or BLOCKED
target_question, target, target_element	the labelled target question (click_target, ...), its gold key and the element's text
source, collection	origin dataset and the collection it was built in (m2w-v2, teacher-v3-glm, dagger-g1, ...)
website, task_id, episode_id	site, task and episode (or trajectory) ids
actor	onpolicy_glm only: who acted at this step, student or teacher
teacher, reason	teacher rows: teacher model and its stated reason

bench adds bench_subset, the config each row comes from. live_tasks has id, url, goal, source.

A request looks like this (shortened):

{
  "model": "wev-latest",
  "state": {
    "page": {"url": "...", "title": "...", "text": "visible page text"},
    "elements": [{"index": "1", "role": "button", "label": "Search", "operations": ["CLICK"]}],
    "recent_actions": [{"action": "...", "kind": "click", "text": null, "page_changed": true}]
  },
  "questions": {
    "operation": {"type": "choice",
                  "instructions": {"goal": "...", "rules": ["..."]},
                  "criteria": {"CLICK": "...", "TYPE_TEXT": "...", "DONE": "...", "BLOCKED": "..."}},
    "click_target": {"type": "choice",
                     "instructions": {"goal": "...", "operation": "CLICK", "rules": ["..."]},
                     "criteria": {"1": {"element": "[1] Search", "current_value": "", "role": "button"}}}
  }
}

A step asks for the operation and, for each operation that needs one, a target (click_target, type_text_target or select_target). A step is answered correctly when the operation and, if the operation takes one, its target both match. Targets are chosen among the candidates listed in the state, not among every element on the page. To train with the wev package, pass JSONL folders to wev train --data; scripts/build_hf_dataset.py documents the mapping.

How the configs were built

mind2web. Each recorded step becomes one request containing the page's visible text, a sample of 8–40 candidate elements that includes the gold element, the task and the preceding actions. Splits are by website, so every test website is unseen in training. Mind2Web has no DONE, WAIT, SCROLL or BLOCKED steps.

nnetnav_audited. NNetNav supplies the stopping (DONE), giving-up (BLOCKED) and scrolling steps that Mind2Web lacks. Its trajectories come from unsupervised exploration, so its stopping labels are noisy. An LLM judge (DeepSeek-V4.1-Flash) read the goal and the state at every step and decided whether the goal was already achieved. On the training split it rejected 592 of 1,898 DONE labels (31%) and found the goal already achieved at 2,281 of the 10,102 other steps (23%). Those steps were relabelled DONE and the rejected DONE steps were dropped. The validation and test splits were audited the same way.

webchain. WebChain v2 releases screenshots and multimodal SFT conversations, but its action metadata also links, for most steps, a pre-action HTML snapshot and an accessibility tree together with the CSS selector of the element acted on. wev/webchain.py resolves that selector in the snapshot (unique matches only, climbing from e.g. a <span> to its <button>), samples 8–40 negatives from the snapshot's other visible interactive elements, and renders the request exactly as for Mind2Web. Clicks, hovers and double clicks become CLICK, typing and pasting TYPE_TEXT. 4,000 of WebChain's 31,445 English-goal trajectories were converted (a fixed hash order); pages not in English, steps without a snapshot, unresolvable selectors and custom (non-<select>) dropdowns were skipped. Splits are by host (175 / 15 / 16 hosts). Like Mind2Web, it has no DONE, WAIT, SCROLL or BLOCKED steps.

teacher_qwen3max, teacher_glm. A prompted LLM served as the System One of jev-ultrafast on live websites: qwen3-max for the two collections in teacher_qwen3max (the second ran on the tasks the first had not solved), and GLM-5.3-Flash, served through Alibaba Bailian, for teacher_glm. The prompt forbade signing in, registering, buying, booking, posting, messaging and submitting personal information, and asked for BLOCKED on CAPTCHAs, unusual-traffic pages and login walls. Every request and the teacher's choice were logged. An LLM judge then read the final page of each episode, and the kept rows are:

• every decision of an episode whose DONE the judge confirmed;
• every decision up to a BLOCKED on a page that blocked the agent;
• every decision but the last of an episode cut short by the browser harness.

Episodes that looped or exhausted their budget were dropped, as were repeated states. The judge rejected 155 of qwen3-max's 431 DONE claims (36%) and 127 of GLM-5.3-Flash's 530 (24%).

onpolicy_glm. Trained only on teacher episodes, a student never sees the states its own mistakes lead to. This config was collected DAgger-style with a mixed policy: at every step both a wev student and the GLM-5.3-Flash teacher answered the request, one of them chosen at random with equal probability acted (actor: 2,894 student and 2,896 teacher steps in train), and the teacher's choice was always recorded as the label. An episode's ending was judged from the actions actually executed: the judge-confirmed final state is labelled DONE, earlier teacher DONE labels (unverified) are dropped, and a BLOCKED ending is kept only when the teacher agrees. The judge rejected 135 of 587 DONE endings (23%). Otherwise the filters are those of the teacher configs.

Splits of the teacher configs. Episodes are split by a hash of the task id, so a task is in the same split in every collection. A held-out task becomes test only when neither its task id nor its goal occurs in any train split of any config, and validation otherwise. NNetNav-sourced tasks take their goals from NNetNav's train split, so the teacher test splits hold only the Google Flights and Wikipedia tasks: 323 steps from 8 tasks.

live_tasks. Goals paired with start URLs: NNetNav goals on their live sites, Wikipedia look-ups and Google Flights searches. The 153 test tasks are the end-to-end suite used to evaluate the wev models; none of them appears in a teacher config. Checked by task id and by goal against every train split: one eval goal ("Find hotels in San Francisco." on booking.com) also occurs at one step of nnetnav_audited train, because NNetNav repeats some goals across its own splits. The 123 NNetNav eval tasks come from NNetNav's test split, so their goals appear in nnetnav_audited test (555 steps); that is by construction, not a leak.

wev-bench

bench is a frozen test set for browser decision models: 2,346 steps, each a request a System One must answer.

Subset	Steps	What it tests
mind2web test	873	clicks, typing and dropdowns on 8 websites unseen in training
nnetnav_audited test	1,150	stopping (DONE 382), giving up (BLOCKED 29) and scrolling on live sites
teacher test slices	323	qwen3-max (57), GLM (147) and on-policy (119) steps on tasks in no train split

Two Mind2Web test requests are left out because one question exceeds the API's 255-option limit.

Metrics. A step succeeds when the operation and, if the step has one, the target are right. wev-bench reports step success, operation accuracy, target accuracy, the premature-DONE rate (steps whose gold operation is not DONE that the model answered DONE: an agent that stops early abandons its task), DONE recall, top-label ECE (10 bins; on the probability the model gives its own choice) and p50/p95 latency per request, overall and per subset. Conventions follow JevBench: failed or malformed answers count as wrong, and an answer without a distribution is scored as one-hot and flagged.

Run it against any endpoint that speaks POST /v1/systemone:

wev serve --model alanhuangya/OpenJev-4B --port 8009        # or any System One server
python scripts/wev_bench.py --url http://127.0.0.1:8009/v1/systemone --name my-model \
  --hardware "RTX 5090, bf16" --out results/wev-bench/my-model.json

The script is in the wev repository and needs only the Python standard library and pyarrow. The result is a JSON file in JevBench's per-system layout (key, display, open, licence, has_distribution, a browser lane block with by_family, speed_block, ...), plus the dataset revision and the bench file's sha256, so it can be submitted to a leaderboard as is.

System One	Step success	Operation	Target	Premature DONE	ECE	p50 / p95 latency
wev-4b	68.0	81.3	79.0	5.0	0.051	166 / 409 ms

wev-4b was served with wev serve on one RTX 5090 (bf16) and queried serially from the same host. Its training included mind2web, nnetnav_audited and teacher_qwen3max train splits, so it is in-distribution there.

Personal and sensitive information

All step rows were scanned for e-mail addresses, phone numbers, card-like and SSN-like numbers, API-key, JWT and private-key shapes, credential-like URL parameters, values in password fields, and personal names typed into fields. Most matches are public business contacts in page text and element labels (phone numbers on 2,510 rows, e-mail addresses on 1,525). Text an agent typed: made-up contact details dictated by Mind2Web and NNetNav goals (e.g. an e-mail on gmail.com, example.com or sample.com), names typed into name fields (23 steps, plus 75 where the goal supplies the name), and values typed into password fields by NNetNav's exploring agent (17 steps) and a 4-digit gift-card PIN in Mind2Web (1). URLs: short-lived signed tokens (JWT-shaped edgeauth/OAuth state values, session ids; 163 rows). Page text: one Hugging Face token in a code sample on a public model page (2 rows) and one store loyalty number (2 rows). No credentials were entered by the teacher, which was instructed never to sign in.

Applied in this release (scripts/redact.py, run by scripts/build_hf_dataset.py): typed e-mail addresses, phone numbers and personal names not dictated by the goal became user@example.com, 555-0100 and Alex Doe; values in password fields, API keys and tokens, and credential-like URL parameters became <redacted>; other e-mail addresses in page text and labels became <email>; the 3 rows with card- or SSN-like numbers were dropped. Published business phone numbers on public pages were kept. Separately, 86 teacher and on-policy steps labelled DONE on a page that had not loaded (no elements, almost no text) were dropped.

Terms

Source	License
Mind2Web-derived rows	CC BY 4.0
NNetNav-derived rows	Apache-2.0
WebChain-derived rows	CC BY 4.0
teacher, on-policy rows and live tasks	see below

Attribute the original datasets (see Citation). The teacher and on-policy configs and the live tasks contain text captured from public websites, which remains subject to those sites' terms. The labels in teacher_qwen3max are outputs of qwen3-max, and those in teacher_glm and onpolicy_glm are outputs of GLM-5.3-Flash (served through Alibaba Bailian); they are subject to their providers' terms. The DONE verdicts and NNetNav relabelling come from an LLM judge (DeepSeek-V4.1-Flash). Treat these configs as research data and check the applicable terms before any commercial use. WebChain's page snapshots were captured by its annotators on live websites and remain subject to those websites' terms; its card asks users not to attempt to recover personal information.

Citation

Jun Huang and Xin Ren contributed equally (University of Electronic Science and Technology of China).

@misc{huang2026wev,
  title     = {wev: Distilling LLM Browser Agents into Open, Local System-One Decision Models},
  author    = {Huang, Jun and Ren, Xin},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22941164},
  url       = {https://doi.org/10.5281/zenodo.22941164}
}

If you use the converted configs, please also cite their sources:

@inproceedings{deng2023mind2web,
  title     = {Mind2Web: Towards a Generalist Agent for the Web},
  author    = {Xiang Deng and Yu Gu and Boyuan Zheng and Shijie Chen and Samuel Stevens and Boshi Wang and Huan Sun and Yu Su},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}
@misc{murty2024nnetnav,
  title         = {NNetNav: Unsupervised Learning of Browser Agents Through Environment Interaction in the Wild},
  author        = {Shikhar Murty and Hao Zhu and Dzmitry Bahdanau and Christopher D. Manning},
  year          = {2024},
  eprint        = {2410.02907},
  archivePrefix = {arXiv}
}
@misc{webchain2026,
  title         = {WebChain: A Large-Scale Human-Annotated Dataset of Real-World Web Interaction Traces},
  author        = {Sicheng Fan and Rui Wan and Yifei Leng and Gaoning Liang and Li Ling and Yanyi Shang and Dehan Kong},
  year          = {2026},
  eprint        = {2603.05295},
  archivePrefix = {arXiv}
}