DATA_FORMAT.md

Data Format

The principal dataset files use JSON Lines format compressed with gzip.

Each decompressed line is one complete JSON object.

Page-level export

File:

data/ultra_pages_v2_redactions.jsonl.gz

Each record represents one scanned PNG page.

Example records

Example document, page, and redacted page records can be found in examples/

Unreliable page records

There are unreliable records that couldn't be transcribed retained in the dataset. Here's an example of an unreliable record:

  "record_version": "ultra-page-v2-redactions",
  "record_id": "0000017354_0030",
  "doc_id": "0000017354",
  "page": 30,
  "source": {
    "filename": "0000017354_0030_1248_1575.png",
    "width": 1248,
    "height": 1575
  },
  "text": "T Sr t o e e t t tee e t e ...",
  "diplomatic_text": "T Sr t o e e t t tee e t e ...",
  "redactions": [],
  "blocks": [],
  "quality": {
    "character_count": 276,
    "word_count": 83,
    "block_count": 0,
    "empty": false,
    "needs_review": true,
    "review_reasons": [
      "severely_corrupted_ocr",
      "no_reliable_sentence_structure"
    ]
  },
  "processing": {
    "engine": "MinerU",
    "mineru_version": "3.4.4",
    "backend": "pipeline"
  }
}``` 

### Page identifiers

- `record_id`  
  Stable page identifier derived from the MORI document ID and page number.

- `doc_id`  
  Zero-padded MORI document identifier.

- `page`  
  Integer page number within the document.

### Source image

- `source.filename`  
  Filename of the corresponding PNG in the IPFS archive.

- `source.width` and `source.height`  
  Source-image dimensions in pixels.

- `source.size_bytes`  
  Original PNG file size, when recorded.

### Text

- `text`  
  MinerU-derived page transcription assembled in reading order.

- `diplomatic_text`  
  The same transcription with experimental redaction markers represented
  separately beneath a `[REDACTIONS]` heading.

### Layout blocks

- `blocks`  
  Reading-order content blocks derived from MinerU output.

- `blocks[].order`  
  Block position in reading order.

- `blocks[].type`  
  MinerU content category, such as text, table, image, or equation.

- `blocks[].bbox`  
  Pixel-space bounding box represented as `[x1, y1, x2, y2]`.

### Quality metadata

- `quality.character_count`  
  Number of characters in the assembled transcription.

- `quality.word_count`  
  Whitespace-based word count.

- `quality.block_count`  
  Number of normalized content blocks.

- `quality.empty`  
  Whether the page has no extracted text.

- `quality.needs_review`  
  Whether automated checks identified a reason for human review.

- `quality.review_reasons`  
  Machine-generated review flags.

### Processing metadata

- `processing.engine`  
  OCR and document-analysis engine.

- `processing.mineru_version`  
  MinerU version used for the export.

- `processing.backend`  
  MinerU backend used.

- `processing.content_list_path`  
  Relative path to the corresponding MinerU content-list artifact.
  
## Document-level fields

- `doc_id`  
  MORI document identifier shared by all included pages.

- `page_count`  
  Number of page records grouped into the document.

- `page_numbers`  
  Page numbers in document order.

- `page_record_ids`  
  Stable page identifiers corresponding to the page-level export.

- `source_filenames`  
  Original PNG filenames in page order.

- `text`  
  Combined transcription with explicit `[PAGE N]` separators.

- `quality.pages_needing_review`  
  Number of pages carrying review flags.

- `quality.empty_pages`  
  Number of pages without extracted transcription text.  
  
## Experimental redaction annotations

The `redactions` array contains probable dark redaction regions detected
from the original page image.

These annotations are heuristic and should not be interpreted as verified
reconstructions of concealed content.

- `marker`  
  Generated textual marker such as `(redacted_12)`.

- `estimated_character_spaces`  
  Estimated number of obscured character spaces based on the region width
  and estimated local character pitch.

- `estimate_range`  
  Approximate lower and upper bounds for the character-space estimate.

- `bbox_pixels`  
  Redaction-region coordinates as `[x1, y1, x2, y2]`.

- `width_px`, `height_px`  
  Region dimensions in source-image pixels.

- `fill_ratio`  
  Estimated proportion of the detected bounding rectangle occupied by dark pixels.

- `aspect_ratio`  
  Region width divided by height.

- `character_pitch_px`  
  Estimated average pixel width per nearby character.

- `pitch_method`  
  Method used to estimate character pitch, such as `nearby_text`,
  `page_median`, or `default`.

- `confidence`  
  Heuristic confidence score, not a calibrated probability.

- `classification`  
  Current label, normally `probable_redaction`.