Release Notes

Changelog

Every version of DocSet Generator, and what changed. Updates are free for all existing customers — re-download the latest build from your original purchase link anytime.

v1.5.0 2026-07-11 Latest

A canonical ground-truth model for all 33 types, ML-ready COCO / LayoutLM / FUNSD / DocVQA exports, reproducible dataset recipes and splits, and Linux (AppImage + Flatpak) builds.

ML-Ready Dataset Exports

  • COCO word-detection exports using rendered-image pixel coordinates, with box scaling fixed to match the rendered image.
  • LayoutLM exports — token, label, and normalized-box JSONL.
  • FUNSD exports with question, answer, header, and other roles plus semantic links.
  • DocVQA exports — questions derived from recorded model fields with answer word/page references.
  • Clean page images for visual-model training, and native-format annotation renditions for .docx, .xlsx, .eml, .csv, and .txt.
  • Deterministic, document-type-balanced train / validation / test splits with a configurable split seed and percentages. The GUI Test percentage is automatic so splits always total 100%.
  • Export reference loaders re-read and validate every generated export.

Canonical Ground Truth & Annotations

  • Pre-render DocumentModel for all 33 document types. Captures ordered titles, headings, paragraphs, list items, and tables — including grids and merged-cell row/column spans — before OCR corruption or rendering, making the pristine model the source of truth instead of recovering it from rendered files.
  • A ground-truth .txt sidecar is written for every generated document; short text also remains inline in the manifest.
  • Rich annotation schema 2.1: model-aligned semantic blocks with word IDs and per-page geometry, explicit line and layout-block boxes, word → line → block parent relationships, and table/row/cell records with geometry.
  • Source-to-rendered token mappings for degraded OCR text, plus expected source text and rendered/degraded text at word level.
  • Generator-recorded, model-derived entities with geometry and key/value pairs (regex entity detection retained as a fallback only), with stricter normalized-box and parent-reference validation.

Reproducibility & Manifests

  • Optional fixed reference date alongside the generation seed, plus derived per-document seeds.
  • Byte-identical seeded PDF generation on the same platform and dependency versions; PDF metadata timestamps and trailer IDs are normalized deterministically (trailer-ID fix keeps xref tables valid for strict PDF readers).
  • Replayable dataset_recipe.json files, with GUI recipe import to review and replay saved settings.
  • Deterministic SHA-256 checksums and archive ordering.
  • Manifest expanded to schema 2.0 — source/rendered/embedded character counts, document checksum, derived seed, reference date, split, locale, generator version, and annotation path. A quality_report.json adds per-type completeness metrics and failures.
  • Bundled JSON Schemas for manifests and annotations, with built-in validation (no extra dependency) shipped in the Windows, AppImage, and Flatpak releases.
  • Prior manifest records are preserved when generating additional documents into an existing dataset folder.

GUI, CLI & Correctness

  • Added COCO, LayoutLM, FUNSD, and DocVQA export controls, split seed and train/validation/test controls, schema validation, Bates prefix/start, PDF metadata author/date overrides, manifest/annotation controls, recipe import, and reference-date controls to the GUI — with matching configurable exports, splits, validation, and recipe replay in the CLI.
  • Incompatible export/validation requests are rejected when manifest generation is disabled; Bates-enabled GUI jobs run sequentially so page numbering stays unique.
  • Reconciliation contracts across every registered type: every type emits a non-empty ordered model, PDF and native model text is complete and does not invent content, and block alignment rejects weak partial matches instead of assigning unrelated geometry.
  • Fixed long CSV ground truth so every delivered row is recorded (with full-row regression coverage for CSV and XLSX), and fixed nested tables in fax cells to contain real cell text.

Performance & Packaging

  • Automatic parallel generation in the GUI with workload-aware worker limits, clean-reference isolation and cleanup, and high-fidelity LibreOffice rendering for native formats when available (with a deterministic PDF fallback).
  • Linux AppImage and Flatpak builds alongside Windows, with reproducible build scripts, release-bundle scrubbing, and bundled Tcl/Tk and JSON Schemas.
  • A shareable annotated v1.5 sample-dataset script, and automated coverage expanded from 33 tests in v1.1 to 104 tests plus native-table subtests.

Internal

  • Product and package version bumped to 1.5.0; Windows executable metadata and bundle naming updated; Linux AppImage and Flatpak build defaults set to v1.5.
v1.1.0 2026-07-03

Ground-truth benchmarking, reproducible datasets, a new document type, and full CLI parity with the GUI.

New Features

  • Ground-truth manifest for OCR benchmarking. Every generation run (GUI and CLI) now writes manifest.jsonl to the output folder. Each entry records the document's intended clean text — captured before any flattening or text-layer corruption — alongside its file path, document type, seed, page count, and the exact corruption settings used. Text over 2,000 characters is stored in sidecar files under ground_truth\. This lets OCR pipelines compare observed output against known-correct text side by side. CLI opt-out: --no-manifest.
  • Seeded, reproducible generation. Pass --seed N (or set a seed in the GUI) to regenerate identical document content with stable, deterministic filenames. Rerunning the same seed at quality 100 produces clean twins of degraded documents.
  • Mass CC Email document type — emails with large visible recipient lists in the To/CC headers.
  • Full CLI option parity with the GUI. New flags: --quality, --long-mode, --watermark, --bates, --bates-prefix, --bates-start, --metadata-author, --metadata-date.

Content & Realism Improvements

  • Email and memo bodies are now contextual — the body language matches the document's subject and type (e.g., mass CC emails use distribution-list phrasing).
  • Invoices and receipts use explicit, realistic sales-tax rates with correct line-item math.
  • Stronger fake-data safety guarantees (zero real PII): email addresses use reserved .example domains, phone numbers stay in the reserved 555-01xx range, and SSNs are structurally invalid on purpose — all enforced by tests.
  • Refined output across many generators: checks, financial statements, government forms, patents, personal information sheets, receipts, transmittals, Word documents, table/list documents, and .eml files.
  • Bates stamping, watermark, and simulated-redaction overlays refined.
  • Batch dataset packager: mixed-type batches now ship with a metadata.csv load file containing page-accurate Bates ranges, custodian, author, and document dates, packaged into a single zip.

Internal

  • All generator registration consolidated into a single canonical registry shared by the CLI, GUI, and batch tools.
  • Test suite expanded to 33 tests; Windows build metadata bumped to 1.1.0.
v1.0.0 2026-04-04

Initial release.

Initial Release

  • 28 PDF document types plus 5 native formats (.docx, .xlsx, .eml, .csv, .txt).
  • Windows GUI and CLI.
  • OCR quality degradation slider, image-only (no-text) PDFs, and corrupt text layers.
  • Long-document mode, watermarks, and Bates stamping.

// Already a customer? Re-download the latest build
from your original purchase link — updates are free.

Purchase — $199 →