Datasets:

Lexmount
/

LexBench-Web-Fetch

pretty_name: LexBench-Web-Fetch
license: apache-2.0
tags:
  - benchmark
  - evaluation
  - browser-eval
  - web-fetch
  - web-scraping
  - web-content-extraction
  - structured-data-extraction
  - document-extraction
  - resource-discovery
  - tool-evaluation
task_categories:
  - other
source_datasets:
  - original
language:
  - ar
  - de
  - en
  - es
  - fr
  - it
  - ja
  - ko
  - la
  - pl
  - pt
  - ru
  - zh
configs:
  - config_name: general-web
    data_files:
      - split: test
        path: data/general-web.jsonl
  - config_name: dynamic-sites
    data_files:
      - split: test
        path: data/dynamic-sites.jsonl
  - config_name: popular-sites
    data_files:
      - split: test
        path: data/popular-sites.jsonl
  - config_name: access-restrictions
    data_files:
      - split: test
        path: data/access-restrictions.jsonl
  - config_name: embedded-documents
    data_files:
      - split: test
        path: data/embedded-documents.jsonl
  - config_name: media-information
    data_files:
      - split: test
        path: data/media-information.jsonl
  - config_name: direct-media-files
    data_files:
      - split: test
        path: data/direct-media-files.jsonl
  - config_name: articles
    data_files:
      - split: test
        path: data/articles.jsonl
  - config_name: dynamic-examples
    data_files:
      - split: test
        path: data/dynamic-examples.jsonl
  - config_name: challenge-examples
    data_files:
      - split: test
        path: data/challenge-examples.jsonl

LexBench-Web-Fetch

LexBench-Web-Fetch evaluates whether a browser, extraction tool or hosted fetch service returns the content an AI agent needs from a URL. Its 550 task entries cover page bodies, structured records with usable links, resource discovery and file delivery. Each task declares what the returned answer must contain, so the same acceptance rules apply to every participant.

Use it to compare live-site fetching, diagnose incomplete extraction, or test a saved response after changing a parser. Browser Eval provides the scorer, execution adapters, evidence storage and comparison reports. Scoring saved answers requires no browser, provider account or LLM judge.

License. The apache-2.0 license covers the project-authored code, documentation and task requirements. The reference texts included for evaluation under definitions/materials/ are third-party content. Apache-2.0 does not cover them, and they keep their source terms; see THIRD_PARTY.md.

Dataset versions

This snapshot is v2.1.0: 550 tasks. See version history and comparison rules for the 550-task and 430-task editions, version increments and immutable revision selection. The active population is recorded in manifest.json under edition; retained historical definitions do not increase that population.

530-task revision

primary-content-20261103 is one cumulative task-scope revision for a 530-task population. It changes 72 cases against the previous revision: 64 scope corrections, article-009 and article-064, three more Xueqiu articles, and binding-only repairs for live-041, live-063 and live-069. The six Xueqiu tasks use one declared shell rule (xueqiu-shell-v1), and the two Quanta tasks use another (quanta-shell-v1). All other case rules and previously released versions remain unchanged. A single compact rationale file explains this rule set; it is not an alternative version.

The revision is intended for formal evaluations under a separate major task contract, using the fixed 530-task selection and fresh matched runs after adoption. It does not certify the old report or establish that all similar requirements in the catalog have been corrected. The 66 earlier cases have source-context review records, which disclose partial observations and unavailable historical captures. The three later Xueqiu cases have no full-page capture among the bound inputs (see the review notes). Review here means assistant-led source inspection and automated checks, not independent human approval. Raw captures are external, so these checks cannot be replayed using the public repository alone. Independent sampling of the removals is still needed before formal adoption. Each case's full-answer control, counterfactual control and source hashes are recorded under cases[] in the rationale file. Freeze the rules and inputs before collecting outputs. Saved-answer rescoring is reported separately as sensitivity analysis. See the scope and adoption requirements.

530-task selection

The selection retains the earlier 430-task edition and adds all 100 article tasks. It excludes both separately labeled 10-task example groups from the mixed webfetch suite. The dynamic examples have corresponding IDs and URLs in the dedicated dynamic suite, where they remain selected. The challenge examples are excluded as a complete additional group, not because all ten are duplicates: only three share a URL with another selected task. This is a report-aligned cohort choice documented after outcomes were available, not a claim of pre-outcome selection or representative sampling. The 530 selected identities and historical editions are unchanged.

The revision uses the explicit 530-task selection: 130 webfetch tasks and 100 each from dynamic-content, popular-sites, access-restrictions and articles. The excluded 20 webfetch entries remain in the historical frozen files. Matching dynamic-content entries remain selected under their own suite identity.

After exporting a suite with the compatible exporter and --revision primary-content-20261103, generate its runner selection from this dataset checkout:

uv run --frozen python tools/task_population.py --suite webfetch

With the compatible Browser Eval population support, prefer --population /path/to/definitions/populations/primary-content-20261103.json --population-suite webfetch on the certified runner, using a dataset exported with this revision. Repeat for each of the five suites. The runner checks the selection against the exported rule hash and frozen suite identity before calls, then records the population hash, selected IDs and rule revision in the result. A single-suite run is not a complete 530-task evaluation. The helper above remains available for legacy manual --case-ids selection, which does not provide this population binding. The historical default viewer and plain webfetch export remain unchanged.

Benchmark categories

The earlier 430-task edition has seven categories: general web content, dynamic websites, popular websites, access-restricted websites, embedded documents, media information and direct media files. The newer 550-task edition adds article tasks and two separately labeled groups of dynamic/challenge examples. Categories describe the task and its required output.

Viewer group	Source suite	Tasks	Scope
`general-web`	`webfetch`	100	Page bodies, documentation and structured records; original categories retain finer detail
`dynamic-sites`	`dynamic-content`	100	Content loaded after the initial response, including asynchronous requests and client-rendered records
`popular-sites`	`popular-sites`	100	Task-relevant stable content and qualifying named records or links on widely used websites
`access-restrictions`	`access-restrictions`	100	Target content at URLs with previously observed access restrictions
`embedded-documents`	`webfetch`	10	Complete document text, including embedded documents, tables and appendices
`media-information`	`webfetch`	17	Requested image and video resources, links or metadata on webpages; media understanding is outside the task
`direct-media-files`	`webfetch`	3	Complete delivered image and video files under the declared file contracts
`articles`	`articles`	100	Article bodies, titles, requested initial comments and post collections
`dynamic-examples`	`webfetch`	10	Additional dynamic-content tasks from the original mixed suite; their declared acceptance contracts remain separate
`challenge-examples`	`webfetch`	10	Additional access-challenge tasks from the original mixed suite; their declared acceptance contracts remain separate

The 100 general-web tasks are the baseline called “static websites” in the published comparison; that name does not mean every target is independent of JavaScript. Media-information tasks request links or page metadata; direct-media tasks require original file bytes. Images and videos can appear in either category. An access-restriction task tests delivery of its target content; a past restriction does not guarantee a challenge on every later request.

The table above identifies this edition's active population. The 550-task edition includes the 100 article tasks and 20 dynamic/challenge examples from the original mixed suite; the 430-task edition excludes them. Keep the example groups' distinct acceptance contracts visible. Definitions retained for reproduction do not add samples to an edition unless its viewer and population manifest select them.

What counts as a pass

A method receives the task URL and prompt. Its response is checked against the task's declared delivery requirements:

Delivery	Checks selected by the task
Body	Required facts and stable text, reference-body coverage, required regions and article endings
Structure	Qualifying records, names and destination URLs, fields owned by the same record, tables or document sections
Resources	Returned resource URLs, permitted variants, and target-only output requirements where declared
Files	Delivered bytes, declared file type and integrity requirements

Contracts can combine these checks. A successful request, nonempty answer or long response alone does not establish task completion. The scorer keeps binary acceptance separate from text-quality diagnostics. Missing reference coverage remains unavailable. Exact requirements and thresholds are in each task; the metrics guide explains their interpretation.

Web Fetch runs call the tested tools directly. An agent does not choose browser actions during execution. Preserve the requested initial-load scope: clicking, logging in, expanding hidden content or paginating changes the work unless a task explicitly permits it. In the optional article extension, a qualified displayed excerpt is scored as that excerpt, and changing feeds use their own post-collection contracts.

A published use of the benchmark

The Web Fetch benchmark comparison evaluates five browsers on these 430 task IDs under its frozen contract edition, and separately on WebMainBench's 545 saved pages. It uses the same URLs and task requirements for each browser, a 45-second task limit, and at most one retry for temporary network errors. The article reports a 96.7% request-success rate but a 54.4% task-pass rate for Chromium, illustrating how returned pages can still miss the requested content. It reports an 89.8% task-pass rate for Moli on that 430-case population.

Those results describe the article's versions, interfaces and execution conditions. They are an example of using per-task evidence to distinguish access failures from extraction failures, and are separate from scores on the current catalog. The article also explains why live fetching and extraction from fixed HTML answer different evaluation questions.

Browse and load the data

Each viewer group has one test split. With the Hugging Face datasets library installed:

import json
from datasets import load_dataset

tasks = load_dataset(
    "Lexmount/LexBench-Web-Fetch",
    "general-web",
    split="test",
    revision="<full-Hugging-Face-commit>",
)
task = json.loads(tasks[0]["task"])
print(task["url"], task["task_prompt"])

Replace the revision with the full dataset commit you intend to use. Select another group by its viewer name in the table above.

Column	Meaning
`id`	Task identifier within its source suite
`input`	Source URL
`benchmark_category`	Task category shared with Browser Eval's report grouping
`source_suite`	Original scorer suite ID, retained independently of the benchmark category
`track`	Original source-group value from the task definition
`category`	Original finer category, preserved exactly as authored
`family`	Declared delivery kind, when present
`language`	Target or reference-content language, taken from source annotations or reviewed frozen references; `mul` marks multiple languages and `zxx` marks URL-only or binary targets
`task`	JSON serialization of the public task definition, including its prompt, acceptance requirements and reference-text hash bindings

The viewer supports inspection and integration. Viewer rows show hashes in place of reference texts. For the 530-task revision, the scorable task records with their reference texts are in definitions/materials/primary-content-20261103/ (see below).

definitions/ contains the public definitions, catalogs and input manifests. definitions.tar.gz packages the same files for installation. manifest.json records source provenance and file hashes. Frozen catalogs retain their original dataset IDs and task languages, including historical method names inside some IDs; the same acceptance rules still apply to every method.

Validate, score and run

From a Browser Eval checkout with uv installed:

git clone https://github.com/lexmount/browser-eval.git
cd browser-eval
uv sync --frozen
uv run --frozen python -m harness.benchmarks.data
uv run --frozen lexbench-webfetch list
uv run --frozen lexbench-webfetch validate

Browser Eval downloads the dataset revision selected by its data lock and verifies the archive and file hashes. Public downloads require no access token. The benchmark integration guide covers revision selection and offline caching.

Score the 530-task revision from this download

definitions/materials/primary-content-20261103/<suite>/dataset.json contains the complete scorable tasks, with primary-content-20261103 already applied, for webfetch, dynamic-content, popular-sites, access-restrictions and articles. Together they cover all 530 selected tasks. Each directory's source-inputs.json and dataset-catalog.json bind the file by SHA-256. The selection's frozen_task_bindings[<suite>].exported_cases_sha256 binds its resolved cases.

Each file is the frozen dataset bound by the suite's definitions/evaluation/.../source-inputs.json. Its hashes were verified, and it reproduces every public task row. The Browser Eval revision code then applied primary-content-20261103, and the result is serialized the way export writes it. These files are ready to score as published, with the revision already applied. Re-running lexbench-webfetch export --revision primary-content-20261103 needs each suite's private frozen parent inputs, including source trees that are not in this download. Re-applying the revision also needs Browser Eval at 351cb27b or later, because its review inherits unchanged entries from review-550-20261023. Install a suite with Browser Eval's existing importer, then run it with the population:

M=/path/to/definitions/materials/primary-content-20261103/webfetch
uv run --frozen python scripts/import_reference_inputs.py --manifest "$M/source-inputs.json" --source "$M"
uv run --frozen python scripts/run_certified_benchmark.py --dataset "$M/dataset.json" \
  --population /path/to/definitions/populations/primary-content-20261103.json \
  --population-suite webfetch --retrieval-mode native ...

To score one saved answer, write a task object from that dataset.json to a file. Declare the answer representation in a metadata file, because text without a declared representation is read as markup:

python -c 'import json,sys; d=json.load(open(sys.argv[1])); json.dump(next(c for c in d["cases"] if c["id"]=="static-simple"), open("task.json","w"), ensure_ascii=False)' "$M/dataset.json"
echo '{"answer_format": "plain_text"}' > metadata.json
uv run --frozen lexbench-webfetch score --task task.json --response answer.txt --metadata metadata.json

Scoring uses only the task object. The declared-retrieval route and the 20 excluded webfetch examples also need the frozen sources and list-references trees, which this download does not include.

Historical editions

For tasks that need external reference inputs, obtain an approved matching copy and export a verified executable suite:

uv run --frozen lexbench-webfetch export \
  --suite webfetch --source /absolute/path/approved-inputs \
  --output /absolute/path/new-input-directory

The scorer's --suite webfetch exports the mixed 150-task suite, including 20 dynamic/challenge examples. For the 430-task edition, select its 130 general-web, document and media IDs plus the three 100-task suites. For the 550-task edition, select all 150 mixed tasks, those three suites and the 100 article tasks (--suite articles). Export verifies inputs before writing a new directory. See reference inputs and access.

Score a saved response using its complete task definition:

uv run --frozen lexbench-webfetch score \
  --task /absolute/path/task.json \
  --response /absolute/path/answer.txt

Add --metadata for saved structured answer fields and --file for delivered file bytes. The Python API, lexbench_webfetch.evaluation.evaluate_response, applies the same acceptance rules.

After exporting webfetch, run one diagnostic task with the HTTP baseline:

uv run --frozen browser-harness-eval run \
  --dataset /absolute/path/new-input-directory/dataset.json \
  --provider-config configs/providers.quickstart.json --providers http \
  --case-ids static-simple --output /absolute/path/new-evidence

The HTTP baseline does not execute JavaScript. For browser setup and a report you can inspect, follow the runnable local example. Formal comparisons use the certified workflow in the Web Fetch evaluation guide.

Reporting and limitations

Report passed/evaluated counts for the declared population, unavailable results and the selected benchmark and scorer revisions. Retain failed calls and timeouts. State timing boundaries, retry policy, local or remote execution, and conversion steps; keep all-call timing separate from timing conditional on success. Optional independent verification has its own disclosed scope and does not run automatically.

Live sites change with location, time, session state and access policy. The suite emphasizes task-relevant stable content, but a frozen reference can still become outdated. Investigate source drift before changing a verdict. Updating task requirements or scoring code does not retroactively update an old measurement. Record the immutable dataset commit, task/reference hashes and scorer source identity with each experiment. Detailed collection, correction and reporting rules are maintained in the Web Fetch guide.

These are curated workloads, not a representative sample of every website. The article suite's initial-load scope does not measure hidden continuations or interaction-dependent content. Text and link checks do not measure visual fidelity or autonomous browser use. This dataset contains task inputs and requirements; it does not ship provider measurements or establish a standing browser ranking.

Sources, related benchmarks and rights

LexBench-Web-Fetch was developed as the Web Fetch benchmark component of Browser Eval. Its task definitions retain source URLs and reference bindings. Project-authored material uses Apache-2.0. The reference texts included for evaluation keep their original language and their source terms, and the Apache-2.0 license does not cover them. To request removal or correction, open a discussion in this repository's Community tab. Full captures, screenshots, document and media files, private logs and measured provider outputs are excluded. A public URL or content hash is not permission to redistribute the underlying material. Read THIRD_PARTY.md, NOTICE and the notices included with the selected snapshot.

The following projects have separate populations and terms; they are not subsets of this catalog:

WebMainBench, maintained by OpenDataLab, evaluates main-content extraction from saved web pages. Browser Eval's LexBench-WebMainBench dataset records its selected corpus and corrections.
WCXB is the Web Content Extraction Benchmark.
WebArena and WebArena-Verified supply interactive tasks and evaluators used in separate browser workflows.
LexBench-Headless-Browser preserves the task inputs for the upstream Headless runtime, which evaluates browser runtime and automation capabilities.

For search evaluation, use LexBench-Web-Search. Moli is one browser that can be evaluated with Web Fetch tasks, and WebFetch CLI is a client for a hosted fetch service. Report creation and sharing are covered by Browser Eval's reporting documentation.

Citation

Cite Browser Eval and record the exact dataset revision used in your experiment. The canonical project citation is CITATION.cff.

@software{browser_eval_2026,
  author = {{Browser Eval contributors}},
  title = {Browser Eval: Benchmarking Browsers and Web Tools for AI Agents},
  note = {Evaluation harness},
  year = {2026},
  url = {https://github.com/lexmount/browser-eval}
}