CORD (receipts) — reshaped mirror

A reshaped mirror of naver-clova-ix/cord-v1 (Park et al., NAVER Clova, 2019), packaged for one-row-per-receipt ingestion. Upstream stores the receipt as a HuggingFace Image feature (a {bytes, path} struct) plus a ground_truth JSON string; pipelines that can't decode a struct image column (or don't want a full datasets dependency) can't consume it directly. This mirror splits each receipt into an image file + a JSONL annotation line, joinable by filename.

Re-hosted under Heliosoph for ingestion-pipeline stability. The annotations are CORD's, unchanged in content; the images are re-encoded (see below). Pin a revision for reproducible fetches.

Credit: Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, Hwalsuk Lee (CORD); NAVER Clova.

Why a mirror?

1. Struct image column. Upstream's image is a HuggingFace Image feature — a {bytes, path} struct in parquet. Readers without nested-type support can't pull the bytes out.
2. Heavy PNGs. The embedded images are full-resolution PNG photos (2 MB each, ~2 GB for the full set). This mirror re-encodes them to JPEG q90 — visually lossless for OCR, ~6× smaller (210 MB total).

The ground_truth annotations are copied verbatim (parsed from the upstream JSON string into a nested object).

What this repo contains

cord-train-images.zip        # 800 receipt JPEGs (entry name == image_file)
cord-train.jsonl.gz          # 800 lines, one JSON object per receipt
cord-validation-images.zip   # 100 receipts
cord-validation.jsonl.gz     # 100 lines
cord-test-images.zip         # 100 receipts
cord-test.jsonl.gz           # 100 lines

Annotation JSONL — one receipt per line

{
  "image_file": "cord-test-0000.jpg",
  "width": 432,
  "height": 648,
  "ground_truth": {
    "gt_parse":   { "menu": { "nm": "-TICKET CP", "cnt": "2", "price": "60.000" },
                    "sub_total": { "subtotal_price": "60.000", "tax_price": "5.455" },
                    "total": { "total_price": "60.000" } },
    "valid_line": [ { "words": [ { "text": "...", "quad": { "x1": 0, "y1": 0, ... } } ],
                      "category": "menu.nm" } ],
    "meta":       { "image_size": { "width": 432, "height": 648 } },
    "roi":        { "x1": 0, "y1": 0, ... }
  }
}

• image_file is the exact entry name inside the images zip — join on it directly, no filename parsing.
• valid_line is the word-level OCR ground truth: each word's text, quadrilateral box, and semantic category.
• gt_parse is the structured key-information-extraction target: menu lines, sub-totals, totals.

How to use

import gzip, json, zipfile, io
from PIL import Image

with gzip.open("cord-test.jsonl.gz", "rt", encoding="utf-8") as f:
    receipts = [json.loads(line) for line in f]

with zipfile.ZipFile("cord-test-images.zip") as z:
    r = receipts[0]
    img = Image.open(io.BytesIO(z.read(r["image_file"])))
    print(img.size, r["ground_truth"]["gt_parse"])

Dataset specs

	Spec
Receipts	800 train / 100 validation / 100 test
Row granularity	one receipt (image + width/height + full ground_truth)
OCR ground truth	word-level: text + quad box + category (valid_line)
Parse ground truth	structured menu / sub_total / total (gt_parse)
Images	JPEG q90 (re-encoded from upstream full-res PNG)
Domain	Indonesian restaurant receipts (photographed)
Format	images as zip per split; annotations as gzipped JSONL

When to pick CORD

• Receipt OCR: detection + recognition on real photographed receipts (skew, glare, thermal paper).
• Key-information extraction / post-OCR parsing: turn a receipt image into structured menu/total fields — the task CORD was built for.
• Document-AI prototyping: word text + box + category is the shape LayoutLM / Donut-style models consume.

For dense printed pages use DocBank; for scene text in the wild, TextOCR / HierText.

License

CC BY 4.0, as released for CORD. Permits commercial use, modification, and redistribution with attribution to NAVER Clova and the CORD authors.

• Paper: CORD: A Consolidated Receipt Dataset for Post-OCR Parsing
• Upstream: naver-clova-ix/cord-v1