Lexmount/LexBench-Zyte-Product-Extraction

Product extraction benchmark (Zyte)

Product extraction benchmark was published by Zyte in 2021. This Hub repository distributes its 140 archived product pages and original references. Lexmount maintains the distribution and a separately identified annotation supplement; it is not the original dataset author. The original source URL and full upstream commit remain part of every release.

Each task asks an extractor to return price, SKU and stock status from the same frozen HTML. The source URL records provenance; requesting the live website changes the input.

Population and references

Viewer group	Source suite	Tasks	Scope
zyte-products-140	zyte-products-140	140	All 140 upstream frozen product pages, original labels and separately identified source-qualified annotations

The upstream source is pinned to commit cba97d7a8d42aeefd4021bc15d482f297faceaa3. All 140 compressed HTML files and the original ground-truth JSON retain their upstream bytes. The runtime manifest also binds the decompressed HTML, original evaluator and license.

Each viewer record contains the source URL, compressed HTML path, HTML hash, original reference, and separate source_qualified_reference. The latter amends 47 fields on 36 pages following a complete model-assisted source audit. It is an additional annotation view, not a second population or human gold. The original reference remains the default for reproducing the upstream score. Reference corrections retain the old and accepted values, literal source excerpts, locators and reasons. Provider outputs, model judgments and measured scores are not part of this dataset.

The original references contain 134 nonempty price fields, 135 nonempty SKU fields and 140 stock labels, including nine out-of-stock labels. The source-qualified view has 134 price fields, 140 SKU fields and 140 stock labels, including 11 out-of-stock labels. Missing values are meaningful and must not become zero prices or automatic in-stock outputs. The original evaluator's stock convention and identifier matching must be retained when reproducing its results.

Load frozen inputs

Select a full Hugging Face commit and keep it with the evaluated outputs:

import gzip
import json
from pathlib import Path
from datasets import load_dataset
from huggingface_hub import snapshot_download

revision = "YOUR_FULL_HF_COMMIT"
rows = load_dataset(
    "Lexmount/zyte-product-extraction-benchmark", "zyte-products-140",
    split="test", revision=revision,
)
task = json.loads(rows[0]["task"])
snapshot = Path(snapshot_download(
    "Lexmount/zyte-product-extraction-benchmark", repo_type="dataset", revision=revision,
    allow_patterns=["definitions/**", "manifest.json", "LICENSE", "NOTICE", "THIRD_PARTY.md"],
))
html = gzip.decompress((snapshot / "definitions" / task["html_path"]).read_bytes())
reference = task["reference"]

Browser Eval uses the shared immutable dataset cache and data lock:

git clone https://github.com/lexmount/browser-eval
cd browser-eval
git checkout 7f010e7617b6ac6a6df366495480af9ac92e4840
uv sync --frozen --extra web-scraper-eval
make pin-dataset DATASET=web-scraper REV=YOUR_FULL_HF_COMMIT
uv run --frozen python -m harness.benchmarks.data web-scraper
uv run --frozen python -m harness.web_scraper.cli prepare

The download verifies the archive and every file. A verified cache works offline with HF_HUB_OFFLINE=1. Historical upstream predictions are fetched separately only by reproduce; they are not extraction inputs. Preparation does not call any extraction provider or require an API key.

Evaluation

Use the original evaluator's per-field precision, recall and F1 for price, SKU and stock. Report out-of-stock recognition separately because the original population is dominated by in-stock pages. Keep transport failures in the predeclared denominator. A structured JSON response alone does not establish field correctness.

Record latency for every logical result, physical retry time separately, and native usage or actual charges where available. API tiers, input limits and billing units differ. This dataset defines extraction inputs and references; it contains no vendor ranking. Browser Eval's execution guide describes native HTML/URL routes and certification.

License, attribution and revisions

Component	Publisher and license	Retained notice
Original 140 archived pages, ground truth, evaluator, README and upstream extractor	Zyte, MIT, Copyright (c) 2021 Zyte	LICENSE and original upstream license
Lexmount-authored packaging metadata and source-qualified annotation supplement	Lexmount, Apache-2.0	LICENSE-annotations and NOTICE
Content captured from third-party websites	Original website publishers retain their rights	Source URLs and THIRD_PARTY.md

The two Hub license badges describe different components. They do not offer a choice to relicense Zyte's original material or the captured websites. The upstream project released its dataset with an MIT notice; this distribution retains that notice and does not independently establish training or redistribution rights for every website asset. Annotation amendments do not replace the original labels.

Cite the upstream product-extraction benchmark and identify this dataset's full commit and reference view. For tooling, cite Browser Eval. VERSIONING.md describes changes to inputs, references and packaging. No moving branch name identifies an evaluated edition.

LexBench-Web-Fetch and LexBench-WebMainBench cover different retrieval and main-content extraction tasks with separate metrics.

Public HTML snapshots

snapshots.json binds every page ID to its SHA-256 and an immutable public deployment. Native extraction APIs receive fixture_base + '/' + id; the original HTML files remain in this repository.

Original files are authoritative. Publication verifies the declared dataset version and complete input inventory before adding this binding. Changed HTML requires a new deployment and a new HF commit; previous files and deployment URLs remain available. Pin the full HF commit for evaluation and verify page hashes before and after calls. A mismatch stops collection rather than silently replacing the input. The navigation index does not host dataset HTML.