Mlchemist/myjev-browser-decisions

myjev browser decisions

Browser-agent steps from five public web-agent datasets, converted into the exact request format that browser-use/jev-ultrafast sends to TypeSafe's Jev decision model, with gold answers. Each record is one browser step: the page state, the user's goal, recent actions and the Choice questions (which operation; which element for each element operation). It is the training and evaluation data of Mlchemist/myjev-decision-4b; code to rebuild it: github.com/KISHAN8888/myjev.

Files

File	Records	Websites	Sources
data/train.jsonl	21,646	301	Mind2Web 5,428 · NNetNav-live 3,983 · Go-Browse 3,961 · NNetNav-WA 3,550 · WebLINX 3,519 · WebChain 1,205
data/validation.jsonl	627	21	Mind2Web 248 · WebLINX 200 · WebChain 179
data/calibration.jsonl	652	28	WebLINX 256 · Mind2Web 200 · WebChain 196
data/benchmark.jsonl	3,345	84	Mind2Web test_domain 1,466 (52 sites) · WebLINX test_web 1,165 (26) · Mind2Web test_website 714 (10)
data/benchmark_jev_answered_ids.txt	3,239 ids		benchmark steps TypeSafe Jev 1.13 answered (106 were rejected as too large); the published comparison uses these
stats.json			build statistics (operations, sources, tags, sites per split, relabel counts, leakage probe)
extras/bench_audit.jsonl	200		label audit of a benchmark sample (verdict, equivalent keys, reason); input to the benchmark build

data/benchmark.jsonl is frozen: SHA-256 f95b894c4161ad7c059624916e6f7e4db9460bbd16c3b16a485481534436bd37.

Gold operation counts:

Split	CLICK	TYPE_TEXT	SELECT	SCROLL_DOWN	SCROLL_UP	WAIT	DONE	BLOCKED
train	14,485	3,171	479	949	129	409	1,785	239
validation	488	71	5	54	9	0	0	0
calibration	531	66	11	39	5	0	0	0
benchmark	2,515	486	95	217	32	0	0	0
benchmark (Jev-answered)	2,471	471	48	217	32	0	0	0

Median elements per step: train 40, validation 30, calibration 34, benchmark 34.

Record format

One JSON object per line:

• request: the body jev-ultrafast posts to /v1/systemone: model, state (page {url, title, visible text ≤ 6,000 chars}, elements [{index, role, label, value, operations, options?}], recent_actions (last 10)), and questions: operation plus one <op>_target question per available element operation (click_target, type_text_target, select_target), each a choice with criteria (the options) and instructions (goal, jev-ultrafast's verbatim prompt text).
• gold: {"operation": ..., "<op>_target": [keys...]}. Several keys = equally valid targets (all must share the operation). SCROLL/WAIT/DONE/BLOCKED have no target.
• meta: id (<source>:<file>:<line>), source, website, site_key (normalised website used for splits), tags (difficulty: many_elements, duplicate_label_distractor, multi_constraint, dropdown_or_autocomplete, date_widget, long_history, terminal_decision), weight (train loss weight), blocked_proxy, bench_part (benchmark), audit_equivalents, plus source-specific provenance (task/trajectory ids, step index, raw action).

Trimmed example (validation):

{"request": {"model": "jev-latest",
  "state": {"page": {"url": "https://www.kohls.com/", "title": "", "text": "Free shipping\nwith\n$49\npurchase...."},
            "elements": [{"index": "1", "role": "link", "label": "Free shipping with $49 purchase. details", "value": "", "operations": ["CLICK"]}, "..."],
            "recent_actions": []},
  "questions": {
    "operation": {"type": "choice", "criteria": {"CLICK": "Click an element, ...", "TYPE_TEXT": "Enter or replace text ...", "SCROLL_DOWN": "Scroll down", "WAIT": "...", "DONE": "...", "BLOCKED": "..."},
                  "instructions": {"goal": "Find mens hiking shoes under $100 in size 10 ...", "rules": "..."}},
    "click_target": {"type": "choice", "criteria": {"1": {"element": "[1] Free shipping with $49 purchase. details", "current_value": "", "role": "link"}, "...": "..."}, "instructions": {"...": "..."}},
    "type_text_target": {"type": "choice", "criteria": {"5": {"element": "[5] Hi James, what are you looking for today?", "current_value": "", "role": "textbox"}}, "instructions": {"...": "..."}}}},
 "gold": {"operation": "TYPE_TEXT", "type_text_target": ["5"]},
 "meta": {"id": "mind2web:train_sample300:73", "source": "mind2web", "website": "kohls", "site_key": "kohls", "step_index": 0, "tags": ["multi_constraint"]}}

Loading: question and option keys vary per record, so Arrow-based loaders (datasets.load_dataset, the dataset viewer) may infer very wide or failing schemas. Reading the JSONL directly is the reliable path:

import json
from huggingface_hub import hf_hub_download
path = hf_hub_download("Mlchemist/myjev-browser-decisions", "data/benchmark.jsonl", repo_type="dataset")
records = [json.loads(line) for line in open(path, encoding="utf-8")]

How it was built

1. Conversion (myjev/convert/). Mind2Web (train, test_website, test_domain) and WebChain: saved page HTML is rendered offline in headless Chromium and jev-ultrafast's real snapshot.js is run; the gold element is mapped to its snapshot index and scrolled to a seeded random viewport height (so its table position does not leak the answer). WebLINX: BrowserGym accessibility tree plus recorded element boxes, reproducing the on-screen rule; goal = instructor messages so far. NNetNav and Go-Browse (successful trajectories only): accessibility tree → snapshot-shaped rows. Actions map to CLICK / TYPE_TEXT / SELECT / SCROLL_* / WAIT / DONE / BLOCKED; goto/back/tab/press steps and steps whose gold element is not in the snapshot are dropped. Requests are built by a port of jev-ultrafast's model.py, tested byte-identical to upstream.
2. Cleaning and splits (myjev/build_dataset.py). Dedupe; website keys normalised across sources (subdomains = same site); splits are website-disjoint (train, validation, calibration and benchmark share no website); train capped at 600 steps per real website, WebArena ≤ 35% of train, DONE ≤ 8%; page_changed blanked and other source-identifying shortcuts removed. Weights: NNetNav 0.5, BLOCKED proxies (stop [N/A], report_infeasible) 0.3, TYPE_TEXT ×1.25, SELECT ×1.5.
3. Label fix (myjev/relabel.py, train/validation/calibration only). Sources disagreed on "click the field, then type" (labelled CLICK in 27–71% of cases); jev-ultrafast types directly. Dropped: CLICK-on-field followed by TYPE into it, Mind2Web Enter-on-field clicks, and TYPE into a field already holding the text (counts in stats.json → relabel).
4. Benchmark = Mind2Web test_website + test_domain + WebLINX test_web, ≤ 120 steps per website, human labels only. A 200-step manual audit found Mind2Web labels ≈ 95% correct or equivalent and NNetNav ≈ 65%, so NNetNav is excluded from all evaluation splits; equivalent alternatives found by the audit were added as extra gold keys (19 steps). The benchmark was then frozen.

Source datasets and licences

Source	Hub id used	Licence	Labels
Mind2Web	osunlp/Multimodal-Mind2Web	CC-BY-4.0	human
WebLINX	McGill-NLP/weblinx-browsergym	CC-BY-NC-SA-4.0	human
WebChain	webagentlab/webchain	CC-BY-4.0	human
NNetNav (live, WebArena)	stanfordnlp/nnetnav-live, stanfordnlp/nnetnav-wa	Apache-2.0	agent-generated
Go-Browse	apurvaga/go-browse-wa	MIT (repository licence)	agent-generated

This dataset is released under CC-BY-NC-SA-4.0 (non-commercial, share-alike) because it includes WebLINX. Each source's own terms also apply to its records; meta.source identifies them.

Known limitations

• AX-tree vs snapshot.js: ~45% of train (NNetNav, Go-Browse, WebLINX) is built from accessibility trees, not the real snapshot.js; NNetNav/Go-Browse rows cover the whole page (no viewport), so element tables are larger.
• Mind2Web rendered without CSS: layout/visibility can differ from the live site; titles are often empty and URLs synthetic. A simple probe can still identify the source from surface features (~82%).
• Label noise: NNetNav and Go-Browse labels are agent-generated (≈ 25% of NNetNav labels wrong in the audit); they are down-weighted and absent from evaluation splits. BLOCKED labels in train are proxies.
• WebLINX conventions: goals are concatenated chat transcripts, reading is labelled as scrolling (no DONE action), and click-then-type is still labelled CLICK in the benchmark. Systems not trained on WebLINX are penalised on that split; report Mind2Web separately.
• Benchmark has no DONE/BLOCKED gold steps; one gold answer per step except audited equivalents (≈ 11% of steps have other valid answers). Candidate sets are viewport-filtered, so scores are not comparable to published Mind2Web numbers.
• Credentials and personal data: WebLINX goals include the demonstration accounts' login e-mails and passwords that the original WebLINX annotators were given (e.g. test accounts on shopping and notes sites), and page text contains public contact e-mails from the recorded websites. These come unchanged from the public sources; records were not edited. Do not use them to log in to anything.

Citation

If you use this data, please cite the source datasets: Mind2Web (Deng et al., 2023, NeurIPS), WebLINX (Lù, Kasner & Reddy, 2024, ICML), WebChain, NNetNav (Murty et al., 2024), Go-Browse (Gandhi & Neubig, 2025), and this release:

@misc{myjev2026data,
  title  = {myjev browser decisions: web-agent steps in jev-ultrafast request format},
  author = {KISHAN8888},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Mlchemist/myjev-browser-decisions}
}