Lexmount/LexBench-NewsXLM

NewsXLM

Short name: LB-NXLM.

This research distribution contains 189 source-qualified pages sampled from NewsXLM. It retains the original benchmark name and attribution. It is a selected evaluation cohort. Scores from this selection do not reproduce the paper's published results.

NewsXLM, https://doi.org/10.15514/ISPRAS-2026-38(2)-10; original MHTML: https://huggingface.co/datasets/ispras-crawlers/newsxlm-mhtml.

Browser Eval

This dataset is integrated with Browser Eval. The Web Extract guide covers dataset selection, hosted or live inputs, provider selection and certified collection. The upstream benchmark links and original license below remain authoritative for the source data.

git clone https://github.com/lexmount/browser-eval
cd browser-eval
git checkout 7e1acca089f04b2bbae4d10890eb298757271a47
uv sync --frozen --extra web-extract-eval

Inputs and references

definitions.tar.gz contains the source-qualified task definition, original HTML, captured resources, references and provenance. manifest.json binds the complete archive and every file by SHA-256. Original task inputs and reference values are unchanged from the preceding edition. Browser Eval uses the shared immutable dataset cache; no extraction service is called during preparation.

make pin-dataset DATASET=newsxlm REV=YOUR_FULL_HF_COMMIT
uv run --frozen --extra web-extract-eval python -m harness.web_extract.cli \
  --home /absolute/data/newsxlm prepare --dataset newsxlm

The selected input mode is snapshot. Native extraction services visit a separately deployed, version-bound HTML service. The deployment exposes pages and resources only; reference answers remain outside the hosted page service.

This cohort requests annotated title, publication date, author and tags where covered. Article body text is excluded. Missing annotations are unscored. MHTML is restored to HTML with captured assets; scored metadata nodes are checked against the original annotations. UTF-8 declarations and local resource links are repaired for hosting. Original MHTML remains in the private source cache.

Scoring

The comparison uses value-level micro F1, page-hit macro F1 and exact-field-set rate. Original references are retained alongside source-qualified whitespace/entity and source-case normalization. Declared URL fields use URL equivalence. These are evaluation adaptations; this distribution does not claim the original paper's protocol.

Rights

NewsXLM by Bedrin et al., 2026; https://doi.org/10.15514/ISPRAS-2026-38(2)-10. The upstream dataset card does not specify a redistribution license. Preserve original website rights and source attribution. Internal research inputs; public redistribution requires approval.

Website content retains its original rights. This repository is private. Packaging does not broaden the upstream license or grant rights to the original website content.

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. Managed immutable deployments are checked through the whole-dataset manifest before and after calls; external hosts have every page byte-checked when first bound. A mismatch stops collection rather than silently replacing the input. The navigation index does not host dataset HTML.