JEVWeb

JEV-format views of OpenJev-Data: the same browser steps, one folder per config, each step unwrapped into the three parts a decision model actually consumes.

Where OpenJev-Data stores a step as the full POST /v1/systemone request in a single request JSON string plus flat metadata columns, every row here is state / questions / gold. The JSONL files keep the native nesting; the parquet files carry the same objects as JSON strings, because a typed struct would inject null keys the source omits (elements without value/options) into training prompts. Nothing else is added or dropped: every request in the source splits back into these columns unchanged.

Load it

from datasets import load_dataset

m2w      = load_dataset("xiq/jevweb", "mind2web")          # human steps, split by website
nnetnav  = load_dataset("xiq/jevweb", "nnetnav_audited")   # live-web steps with audited DONE labels
webchain = load_dataset("xiq/jevweb", "webchain")          # human steps on 200+ live sites
qwen     = load_dataset("xiq/jevweb", "teacher_qwen3max")  # qwen3-max acting as System One
glm      = load_dataset("xiq/jevweb", "teacher_glm")       # GLM-5.3-Flash acting as System One
onpolicy = load_dataset("xiq/jevweb", "onpolicy_glm")      # student-visited states, GLM labels
bench    = load_dataset("xiq/jevweb", "bench", split="test")           # OpenJev-Bench, frozen
tasks    = load_dataset("xiq/jevweb", "live_tasks", split="test")      # 153 end-to-end eval tasks

Configs

Config	Rows (train / validation / test)	Source
mind2web	5,861 / 586 / 875	Mind2Web
nnetnav_audited	11,408 / 1,140 / 1,150	NNetNav-live, stopping labels audited
webchain	17,876 / 1,069 / 1,589	WebChain v2
teacher_qwen3max	4,781 / 385 / 57	qwen3-max as the System One of jev-ultrafast on live sites
teacher_glm	5,969 / 370 / 147	GLM-5.3-Flash in the same role
onpolicy_glm	5,782 / 275 / 119	DAgger-style: student or teacher acts, GLM teacher labels
bench	2,346 test	OpenJev-Bench: frozen test steps from the configs above
live_tasks	1,402 / – / 153	goals + start URLs, shipped unchanged (see below)

Each split is also available as JSONL next to its parquet (webchain/train.jsonl, same rows, same order).

Columns

Step configs (all but live_tasks):

Column	Meaning
state	JSON string: page (url, title, visible text), elements (the 8–40 candidate elements offered), recent_actions
questions	JSON string: the typed questions for this step — operation, plus a target question (click_target, type_text_target or select_target) for each operation that needs one, each with its instructions and criteria
gold	JSON string: the gold option key per labelled question, e.g. {"operation": "CLICK", "click_target": "7"}

live_tasks keeps the source columns id, url, goal, source: a goal and a start URL, no page state. These rows cannot be expressed as state / questions / gold — the request only exists once an agent has visited the site and the page has been parsed into elements. The 153 test tasks are the held-out end-to-end suite.

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 replay a step against a System One endpoint, rebuild the request from the two objects: {"state": json.loads(row["state"]), "questions": json.loads(row["questions"])}.

Terms

Row provenance, licences, the redaction pass and the auditing of the stopping labels are described in the OpenJev-Data card, which this repo re-expresses rather than replaces. Mind2Web-derived rows are CC BY 4.0, NNetNav-derived rows Apache-2.0, WebChain-derived rows CC BY 4.0; teacher, on-policy and live-task rows contain text captured from public websites and remain subject to those sites' terms and to the labelling models' provider terms. Treat them as research data and check the applicable terms before any commercial use. Please cite OpenJev-Data and the original datasets, as that card does.