WebJev

Training web agents to complete real tasks on the live web

To complete a task on a real website, a web agent must get a long chain of decisions right, from the first page to the final answer: what to do next, and which element to act on. WebJev captures these decisions along complete task trajectories on the live web. It contains 64,122 decisions from 3,858 tasks on 1,443 real-world websites. They cover every stage of a task: searching, navigating, filtering, filling in forms, reading results and deciding when the task is done.

Each instance pairs an agent's observation of the current page with one multiple-choice question. Together the instances cover two complementary skills:

• Action prediction (31,050 instances): choose the next operation among those valid on the current page (3–13 options). The operations are click, type, select, scroll, wait, submit, dismiss, go back, finish and give up.
• Element grounding (33,072 instances): choose the element to act on among up to 137 candidates from the same page (median 32).

Observations follow the format a deployed browser agent sends to its decision model. They include:

• the page's URL, title and visible text;
• the actionable elements in the viewport, with their roles, values and positions;
• the recent action history.

An LLM annotator produced the labels while operating the agent on live websites, and only labels that passed multi-signal verification were kept. The grounding instances are further enriched with same-page hard negatives at several candidate scales.

One WebJev task on spanishdict.com, completed decision by decision: type the query, pick the suggestion, dismiss a pop-up, read the entry and finish. The blue boxes mark the elements the agent acts on.

Highlights

• Whole-task coverage. Decisions come from complete task trajectories on live websites. They span every stage of a task, from the first search to the final DONE.
• Live and diverse. 3,858 tasks on 1,443 real-world websites, from shopping, travel and real estate to education, research, software and sports.
• Two decision skills. 31,050 action-prediction and 33,072 element-grounding instances.
• Deployment-faithful observations. 89% of the instances use the structured observation of a production browser agent; the rest use a textual page rendering that adds format diversity.
• Hard grounding. Grounding questions have a median of 32 candidates and up to 137, with same-page hard negatives and multi-scale candidate sets.
• Long context. The median instance has 4,262 tokens, and 10% of instances have more than 9,730.
• Verified labels. Labels are kept only after multi-signal verification: execution on the live page, re-annotation consistency and adjudication. Independent audits of stratified samples found 98% of annotated labels correct or acceptable.

At a glance
Instances	64,122 (31,050 action prediction, 33,072 element grounding)
Tasks · websites	3,858 · 1,443
Distinct decision points	40,230
Candidates per question	action: 3–13 (median 7); grounding: 2–137 (median 32)
Context length	median 4,262 tokens, 90th percentile 9,730, maximum 16,383 (327.6M in total)
Observation formats	structured agent observation 57,283; textual rendering 6,839
Language	English tasks and instructions; page text mostly English
Collection period	August–September 2026

Element-grounding examples from twelve of the 1,443 websites. The blue box marks the gold element; each caption gives the intended step and the number of candidates.

Task formulation

A web task unfolds as a chain of decisions. At every step the agent observes the page, decides what to do and, for clicks, typing and selects, which element to act on. Each decision yields one instance per question.

One decision on spanishdict.com. After typing "spring" into the search box, the agent must decide what to do next (action prediction) and which of the 28 visible elements to click (element grounding).

Each instance is a tuple (observation, instruction, candidates, answer):

• Observation (context). What the agent sees at this step. The structured form is a JSON object:

  • page: URL, title and visible text;
  • elements: actionable elements of the viewport, each with its index, role, label, current value, supported operations and position;
  • recent_actions: the recent action history and whether each action changed the page.

  The textual form renders the website, the task, the current subgoal, the actions taken so far and the page text as plain text.

• Instruction (question). What is asked. Structured instances carry the user's task, the current subgoal, the date and the agent's decision policy. Textual instances carry a single question.

• Candidates (options). For action prediction, the operations valid on the page. For element grounding, the elements of the page that support the requested operation.

• Answer (gold). The index of the correct candidate.

Action space.

Action	Meaning
CLICK	click a link, button, menu item, suggestion or date
TYPE_TEXT	enter text into an editable field
SELECT	choose a value in a drop-down
SCROLL_DOWN / SCROLL_UP	scroll the page
SCROLL_REGION	scroll inside a scrollable page region (appears as E<n> in the options)
WAIT	wait for the page to update
PRESS_ENTER	submit the focused field
PRESS_ESCAPE	close a menu or pop-up
GO_BACK	return to the previous page
DONE	the goal is visibly satisfied
BLOCKED	no available action can make progress

Model interface. The dataset suits decision models that score options rather than generate free text. We render each instance as follows:

Context:
{context}

Question: {question}
Options:
(A) {option}
(B) {option}
…
Answer: (

• The option order is shuffled.
• Options are labelled A, B, … and then AA, AB, … for large candidate sets.
• The model is trained with cross-entropy over the option labels at the answer position.
• num_tokens reports the length of this rendering under the Qwen3.5 tokenizer.

Data distribution

(a) Action labels of action-prediction instances. (b) Composition by question type. (c) Candidate-set size of element-grounding instances. (d) Context length.

Composition

Task type	Question type	Structured observation	Textual observation	Total
Action prediction	operation	31,050	0	31,050
Element grounding	click_target	25,465	6,197	31,662
Element grounding	type_target	541	604	1,145
Element grounding	select_target	227	38	265
Total		57,283	6,839	64,122

Actions

The action labels follow the natural frequency of decisions during web navigation. Clicks and scrolling dominate, and terminal decisions (DONE, BLOCKED) account for 8.6%.

Action	Instances	Share
CLICK	14,364	46.3%
SCROLL_DOWN	6,135	19.8%
TYPE_TEXT	3,527	11.4%
DONE	2,397	7.7%
GO_BACK	1,553	5.0%
WAIT	1,137	3.7%
SCROLL_UP	713	2.3%
PRESS_ENTER	362	1.2%
SCROLL_REGION	355	1.1%
BLOCKED	259	0.8%
SELECT	226	0.7%
PRESS_ESCAPE	22	0.1%

Grounding difficulty

Grounding questions are hard: the correct element must be chosen among many look-alike candidates of the same page.

• Click targets are the largest group. Their candidate sets are multi-scale (32 / 64 / 128 / all elements), which exposes models to increasing distractor density.
• Select targets often list long option menus.
• Type targets usually choose between a few input fields.

Question type	Instances	Min	Median	90th percentile	Max
operation (action prediction)	31,050	3	7	9	13
click_target	31,662	2	32	86	137
type_target	1,145	2	2	4	12
select_target	265	2	26	81	131
all grounding	33,072	2	32	84	137

Context length

Subset	Instances	Median	90th percentile	Max
Action prediction	31,050	3,056	6,902	16,371
Element grounding	33,072	5,665	11,366	16,383
Structured observations	57,283	4,102	9,332	16,383
Textual observations	6,839	6,102	11,587	16,376
All	64,122	4,262	9,730	16,383

Language

• The tasks, instructions and candidates are in English.
• Page text is predominantly English. In 0.9% of the instances a non-Latin script makes up at least 20% of the observation: CJK 434, Cyrillic 65, Devanagari 53, Arabic 36, Greek 5.
• Localized pages in other Latin-script languages also occur.

Label sources

Label source	Instances	Share	Description
live_annotation	43,890	68.4%	action chosen by the LLM annotator on the live page, kept after verification
trajectory	20,011	31.2%	target taken from a successful agent trajectory and verified
trajectory_corrected	221	0.3%	trajectory target replaced by the reviewer's corrected target

Dataset structure

Configs and splits

Config	Instances	Content
all (default)	64,122	all instances
action_prediction	31,050	next-action questions
element_grounding	33,072	click, type and select target questions

The dataset is released as a single train split. Instances that share a task_id come from the same task, and instances that share an instance_id come from the same decision step: its action and grounding questions, or the same grounding question at different candidate scales. Use task_id to build leakage-free held-out splits.

Data fields

Field	Type	Description
id	string	unique instance id
task_type	string	action_prediction or element_grounding
question_type	string	operation, click_target, type_target or select_target
action	string	action prediction: the gold action (see the action space). Grounding: the operation being grounded
context	string	the observation (model input)
question	string	the instruction and question (model input)
options	list[string]	the candidates (model input)
gold	int32	index of the correct candidate in options
answer	string	options[gold]
num_options	int32	number of candidates
num_tokens	int32	length of the rendered instance (Qwen3.5 tokenizer)
context_format	string	structured or text
website	string	the task's target website
page_url	string, nullable	URL of the observed page (structured observations)
task_source	string	public task collection of the task
label_source	string	live_annotation, trajectory or trajectory_corrected (see Label sources)
task_id	string	task identifier, for grouping
instance_id	string	decision-step identifier, for grouping

Only context, question and options are model inputs, and gold is the target. The remaining fields describe the instance and are meant for analysis and splitting.

Example

{
  "task_type": "element_grounding",
  "question_type": "click_target",
  "action": "CLICK",
  "context": "{\"page\": {\"url\": \"https://www.spanishdict.com/\", \"title\": \"SpanishDictionary.com | English to Spanish Translation, Dictionary, Translator\", \"text\": \"Learn Spanish\\nTranslation\\nConjugation\\n…\"}, \"elements\": [{\"role\": \"link\", \"index\": \"1\", \"label\": \"SpanishDictionary.com Homepage\", \"operations\": [\"CLICK\"], \"position\": {\"x\": 213, \"y\": 0}}, …], \"recent_actions\": [{\"action\": \"Translate Spanish or English\", \"kind\": \"fill\", \"text\": \"spring\", \"page_changed\": true}]}",
  "question": "{\"goal\": \"… The user's original task: On https://www.spanishdict.com: Search for the word 'spring' on SpanishDict and identify two different Spanish translations that represent distinct meanings. …\", \"operation\": \"CLICK\", \"rules\": [\"Advance the user's entire goal from the CURRENT page using one operation. …\"]}",
  "options": ["1: {\"element\": \"[1] SpanishDictionary.com Homepage\", …}", "…", "14: {\"element\": \"[14] spring\", \"role\": \"option\", …}", "15: {\"element\": \"[15] spring break\", \"role\": \"option\", …}", "…"],
  "gold": 12,
  "answer": "14: {\"element\": \"[14] spring\", \"current_value\": \"0\", \"role\": \"option\", \"selected\": \"false\"}",
  "num_options": 28,
  "num_tokens": 3949,
  "context_format": "structured",
  "website": "spanishdict.com",
  "task_source": "MolmoWeb",
  "label_source": "live_annotation"
}

Usage

from datasets import load_dataset

ds = load_dataset("Lexmount/WebJev", "all", split="train")
grounding = load_dataset("Lexmount/WebJev", "element_grounding", split="train")

def render(ex):
    labels = [chr(65 + i) for i in range(26)] + [a + b for a in map(chr, range(65, 91)) for b in map(chr, range(65, 91))]
    options = "".join(f"\n({labels[i]}) {o}" for i, o in enumerate(ex["options"]))
    return f"Context:\n{ex['context']}\n\nQuestion: {ex['question']}\nOptions:{options}\nAnswer: ("

# leakage-free split by task
tasks = sorted(set(ds["task_id"]))
held_out = set(tasks[: len(tasks) // 20])
train = ds.filter(lambda ex: ex["task_id"] not in held_out)
valid = ds.filter(lambda ex: ex["task_id"] in held_out)

Dataset construction

WebJev was built in four stages.

1. Task curation. We build on the task collections of MolmoWeb and WebGym. We keep tasks on real, publicly reachable websites that need no account.

2. Agent-in-the-loop collection. An LLM annotator (DeepSeek-V4.1-Flash) operated a production-grade browser agent on the live websites, in isolated cloud browser sessions.

   • At every step, the agent's observation and questions were recorded exactly as served.
   • The annotator's choice was then executed, and its effect on the page was recorded with it.
   • A successful demonstration of the same task guided the annotator. It is never part of the model input, so each instance must be answerable from the observation alone.

3. Multi-signal verification. Each candidate label was checked against complementary signals:

   • execution evidence and a meaningful page-state change;
   • trajectory-level loop detection;
   • agreement under independent re-annotation with shuffled options and without guidance;
   • adjudication by an independent verifier for terminal and ambiguous decisions.

   A final review restored valid decisions that conservative checks had rejected and removed near-duplicates. About half of all recorded decisions survived verification.

4. Grounding enrichment. Successful steps of agent trajectories on the same websites were converted into additional element-grounding instances.

   • Candidate sets are built from the actionable elements of the same page, with mined hard negatives.
   • Each question is rendered at several candidate scales.
   • An LLM reviewer checks the target, and a blind challenge removes candidates that could also be correct.
   • A textual observation view adds format diversity.

Quality control.

• Independent audits of stratified samples, checked against page screenshots, found 98% of the annotated labels correct or acceptable.
• Every instance fits in 16K tokens without truncation.
• Exact duplicates were removed.

Licensing and access

• Licence. WebJev is released under the Apache License 2.0. You may download it, use it to train and evaluate models, including commercial ones, and redistribute it. Please keep the licence and attribution notice.
• Third-party rights. Page text, URLs and element labels were captured from publicly accessible websites and remain the property of their owners. The task descriptions derive from public task collections under their own licences, and the labels were produced with DeepSeek models under their terms of use. The Apache licence covers Lexmount's contribution (collection, labels, annotations and structure). It grants no rights in third-party content.

Citation

@misc{lexmount2026webjev,
  title        = {WebJev: Training Web Agents to Complete Real Tasks on the Live Web},
  author       = {{Lexmount}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Lexmount/WebJev}}
}

Contact

Lexmount, via the Lexmount organization on Hugging Face.