LexBench-WebMainBench

A reference-corrected edition of OpenDataLab WebMainBench with 545 frozen extraction tasks, used by Browser Eval. Each task supplies frozen HTML and a reference Markdown answer, so extractors can receive the same page bytes. This edition supports comparisons of main-content extraction, HTML-to-Markdown conversion and preservation of code, formulas and tables.

Viewer group	Source suite	Tasks	Scope
corrected-545	corrected-545	545	Frozen HTML with corrected reference Markdown and feature-scope annotations

What changed from upstream

The corrected-545 configuration retains all 545 pages in their original order, with unchanged track_id, URL and input HTML. It uses the upstream 545-page subset with calibrated references; it does not contain the full 7,809-page WebMainBench corpus.

The correction receipt records 22 corrections: 19 feature-scope metadata changes and three replacement reference texts. One replacement removes duplicated hidden editor content; two preserve selected inline code as inline code in the Markdown reference. Read the per-case rationale and source hashes in the correction manifest; the three replacement Markdown files are beside it. Upstream provenance binds the original corpus, the corrected file and the matching correction/scorer revision.

Browser Eval's WebMainBench fork also carries scoring corrections for component applicability and aggregation. Dataset and scorer revisions must be recorded together. Results from this edition need to name the corrected edition and evaluation configuration; they are distinct from the upstream leaderboard.

Data and schema

Resource	Contents
data/corrected-545.jsonl	Dataset-viewer rows: id, input, family, language, task
definitions/WebMainBench_545.jsonl	Complete corrected source records, one JSON object per line
definitions/corrections/	Correction receipt and three replacement references
definitions/upstream.json	Source revision, corpus hashes and correction/scorer revision
manifest.json	Release source commit, file sizes and SHA-256 hashes

The viewer's id is the original track_id; input is the page URL. Parse task as JSON for the full record: track_id, url, html, main_html, convert_main_content, groundtruth_content and meta. Metadata includes the original page language and content-feature annotations. The viewer's language describes the frozen reference text: reviewed labels fill missing annotations and correct demonstrated source errors; mul marks references with substantial content in multiple languages. Original meta.language values remain unchanged inside task. Language-annotation provenance is recorded in manifest.json.

html is the frozen extraction input. groundtruth_content is the corrected evaluation reference. URLs identify the source pages; fetching those live URLs produces a different input from replaying the saved HTML.

Load and evaluate

Choose a full Hugging Face commit from this dataset's revision history and retain it with your results:

import json
from datasets import load_dataset

rows = load_dataset(
    "Lexmount/LexBench-WebMainBench",
    "corrected-545",
    split="test",
    revision="YOUR_FULL_HF_COMMIT",
)
page = json.loads(rows[0]["task"])
html, reference = page["html"], page["groundtruth_content"]

For a raw-file download, definitions.tar.gz contains the same definition bytes as definitions/; verify them against manifest.json. Browser Eval selects its dataset and scorer through the data lock. Its benchmark guide explains the pinned runner and scorer. Execution starts at scripts/run_webmainbench_benchmark.py; scripts/score_webmainbench_results.py scores saved outputs.

The 545-page scorer measures text, code, formula and table edit similarity, plus table structural similarity (TEDS). Report the components, their applicable populations and the selected aggregation. The full upstream corpus's ROUGE evaluation is a separate protocol. Provider outputs and measured scores are not included in this dataset.

License and attribution

OpenDataLab WebMainBench and this edition's project-authored changes use Apache-2.0; see LICENSE, NOTICE and THIRD_PARTY.md. The package includes upstream website HTML and reference text. Website content retains its source rights; the repository license does not establish blanket rights to reuse every captured page for model training.

Please cite the upstream work and identify this corrected dataset revision:

Liu et al. (2025). Dripper: Token-Efficient Main HTML Extraction with a Lightweight LM.

For the execution and reporting harness, cite Browser Eval: Benchmarking Browsers and Web Tools for AI Agents.

Related datasets: LexBench-Web-Fetch covers additional fetch tasks; LexBench-Web-Search and its separate SimpleQA Verified retrieval edition cover search. Their task populations and evaluation contracts are separate from these extraction pages.