Detection fine-tune export

Related: finetune-export.md, detection-finding.md, product outline …/finetune_export.md.

Purpose

Export page images + effective text with approved / rejected detection finding labels as a local dataset for external detector fine-tuning. Transcribe does not train models.

Package layout

detection_finetune_export_<stamp>/
  manifest.json
  samples.jsonl
  images/<page_id>.png
  README.txt

• manifest.json format: transcribe.detection-finetune-export-manifest schema_version 1

• Knobs: include_unreviewed (default false), detector_ids, finding_types, image_mode (copy | hardlink)

JSONL sample (v1)

{
  "id": "<notebook_id>:<finding_id>",
  "image": "images/<page_id>.png",
  "text": "<page effective text>",
  "label": "approved|rejected|unreviewed",
  "finding": {
    "finding_id": "...",
    "detector_id": "...",
    "finding_type": "...",
    "snippets": [],
    "prompt_id": "...",
    "prompt_version": "..."
  },
  "source": {
    "notebook_id": "...",
    "page_id": "..."
  }
}

Unreviewed findings are excluded by default. Filter by detector_id / finding_type via CLI flags or options.

Surfaces

• CLI: transcribe export-detection-finetune <project> [dest]

• UI: Workflow → Export → Detection fine-tune

Privacy

Images leave the machine only if the user copies the export folder. No network upload from this export path.