Datasets:

---

Lexmount

/

LexBench-Web-Fetch

Like 1

Follow

Lexmount 9

Make the 530-task population (definitions/populations/primary-content-20261103.json cases) the active edition v3.0.0: the earlier 430-task population plus 100 article tasks.

The 550-task catalog counted two 10-task example groups from the mixed webfetch suite:

• 10 dynamic-examples have the same IDs and URLs as tasks in the 100-task dynamic-content suite; counting them double-counts those targets. Their viewer config and JSONL are removed.
• 10 challenge-examples (Amazon, JD, Xiaohongshu, Zhihu, Douban, Dianping, Reddit, IMDb, Anthropologie, JustWatch list pages) are an example group. They stay as a separately reported challenge set, outside the counted edition.

Changes:

• manifest.json edition: tag v3.0.0, version 3.0.0, task_count 530, eight configurations, population_sha256 d6914e377dfc11c07627a733b3aa810a2b9fe7adf86e4acf85ad140680ecdd25 (VERSIONING.md method). Viewer metadata drops dynamic-examples and labels challenge-examples as uncounted.
• README/VERSIONING: v3.0.0 (530 tasks), v3.0.0 row, 550 described as the historical v2.x edition; main/full content scopes apply to these 530 tasks.
• tools/test_content_scopes.py: edition is 530 and selects exactly the population cases; challenge set published outside the edition.
• Archive/manifest rebuilt with tools/rebuild_publication.py, source_revision a71b0b0fea2f935fa4217c8a2eb066707687c0f9.

All frozen definitions are unchanged; v1.0.0 and v2.x remain reproducible from their tags. make check: 37 tests OK; rebuild check {"task_count": 530, "archive_members": 65}. Do not tag before explicit release authorization.

Fix commit b3f6b5e0 (review findings):

• The title's "without duplicated examples" is narrower than it reads. The 10 dynamic examples share IDs and URLs with dynamic-content tasks, but their acceptance contracts differ; the edition keeps the dynamic-content contracts. VERSIONING.md v3.0.0 row and README now say this.
• README states that the 530-task selection was documented after outcomes were available: it is a rule-based cohort boundary, not a pre-outcome selection protocol or a representative sample.
• README export guidance now says "historical 550-task (v2.x) edition".

Archive and manifest rebuilt with tools/rebuild_publication.py (source_revision a71b0b0f). definitions.tar.gz sha256 bc1ff8dd5de1b1497483993b47dbe28a51dd532e835e6006a7f4862e10c1692d. make check on the PR head merged with main: 37 tests OK, {"task_count": 530, "archive_members": 65}. Not merged or tagged.