mannycooper/document-review-data

Document Extraction Data / 文档提取数据集

Public data release authorized by the repository owner. Mixed source licensing and provenance remain per-source; no blanket license is asserted.

Current full release: 20261015_cee86d566766. Start with the catalog and collections.

• Title selection: current human core 4000 rows / 3918 concrete labels / 79 skips / 3 parser exclusions, with 3999 parse XML files. The active label and modeling authority remains canonical baselines.json.
• Human expansion: 1000 source documents, 997 bound XML files. Independent human labels remain separate from automatic visual labels on overlapping documents.
• Vertical expansion: 4473 scored and 154 unresolved documents; 4239 selected, 1400 category-balanced and the previously published 1200 report view are overlapping development subsets. Do not sum them or call them blind tests.
• Document structure: SoMark full-document and heading pioneer references, balanced views, alignment evidence and historical heading experiments. Provider references and automatic alignments are not human structure truth.
• Parser/real-device evidence: 440 synthetic cases (438 content identities; 250 expected JSON files), DFX human subset and 280 runs on 140 unique documents.

Historical data is stored in content-deduplicated shards, with every source path, content hash, transformation and part offset indexed. Offline restoration instructions and the restore tool reconstruct package aliases and historical files.

An additional whole-inventory archive retains unselected collection-pool documents, older parser XML, structured experiment outputs, arrays and diagnostic images beyond the named collection manifests. Per-file dispositions distinguish verified archived bytes from files excluded at freeze; labels and parser-version bindings remain unknown unless independently documented. These assets do not enlarge the active training baseline.

Additional audited duplicate paths reference bytes already covered by the main release, including duplicate archive members. Their identity is the audit-time SHA; this index does not claim a fresh read of mutable local files or infer human labels. See the index summary.

Known gap: 564 historical full-document device XML files from task0149 are absent locally and have not been uploaded. Receipts and locations are preserved; first-page XML is not substituted. Missing/failed parses retain explicit status and never become negative labels.

The Viewer reads only explicit display tables; XML/Office/PDF files and historical shards are downloadable assets. Existing Viewer configurations are retained as historical views. Old training split names do not re-establish an independent holdout.

Existing paths and history remain available; no remote deletion is part of this release. Predictions, machine labels, human decisions, programmatic expectations and research features have distinct provenance. Current OOF modeling assets are research evidence, not a deployed production model.

Legacy review-app entry remains manifests/hf_manifest.jsonl.