K10124/japan-ocr-mini-benchmark

Japan OCR Mini Benchmark

Japan OCR Mini Benchmark is a compact, local-first benchmark for testing whether OCR/VLM models can read Japanese receipts and return structured JSON.

It is not just a text-reading demo. It checks whether a model can recover receipt-level fields, item rows, taxes, totals, quantities, discounts, payment methods, and noisy camera-like variants.

領収書 / 明細 / 税 / 合計
Synthetic Japanese receipts with fictional stores, dense line items, and model-ready ground truth.

Latest Status

• Current dataset payload: v0.2.0
• Current model benchmark: v0.4.0 Clean/Noisy LM Studio leaderboard
• Current receipt-generation library: v0.4.2
• Accepted receipt design candidates: 83
• Distinct semantic structures: audit in progress
• Selection review: internal curator review, summarized in reports/reference_generation/v0.4.2/summary.json
• Accepted design report: reports/reference_generation/v0.4.2

Why This Exists

Most OCR examples stop at "can the model read the text?" Japanese receipts are meaner and more interesting:

• item names are short, dense, and often abbreviated
• tax categories can mix 8%, 10%, non-taxable, discounts, and coupons
• totals must agree with item rows
• noisy photos bend, blur, fade, crop, and shadow the paper
• local VLMs often produce plausible JSON that is subtly wrong

This project gives you a small but inspectable target for fast local testing before scaling to a bigger evaluation.

What Is New In v0.4.2

v0.4.2 turns the receipt-generation work into a curated design-candidate library.

• 83 accepted synthetic clean receipt design candidates
• the 83 candidates include semantic structures, layout variants, and branding/logo/typography variants
• audited distinct semantic structure count is still in progress
• Japanese logo-like store names, brush-style headers, seals, dense supermarket receipts, food service receipts, payment/point/coupon layouts, and specialty receipt formats
• accepted / hold / rewrite review workflow for future expansion
• public-safe fictional data policy retained
• real reference images are not copied into the public report

Benchmark Releases

Area	Version	What it contains
Dataset payload	v0.2.0	20 synthetic Japanese receipts with clean/noisy images and ground-truth JSON
LM Studio baseline	v0.3.0	first local multi-model noisy-image benchmark
Operational snapshots	v0.3.1, v0.3.2	additional LM Studio runs and combined rankings
Clean/Noisy leaderboard	v0.4.0	paired clean and noisy benchmark tables
Generator QA	v0.4.1	audited 23-template clean receipt generator snapshot
Design-candidate library	v0.4.2	83 accepted receipt designs for future taxonomy and 100-type generation

Structure Selection Method

The generator now separates three ideas that used to be easy to mix up:

• candidate: a generated receipt design worth reviewing
• distinct structure: a receipt with meaningfully different fields, tax/accounting logic, payment behavior, service flow, or document topology
• visual variant: a layout, logo, store-name, brush-lettering, seal, font, or branding difference

The 100-type target means 100 audited distinct structures. It does not mean 100 images or 100 logo variants.

Leaderboards

The current model leaderboard uses JOMB Core Score v1:

Core Score =
Exact match       * 10%
+ Top-level fields * 25%
+ Item fields      * 50%
+ Item count       * 15%

Read the full ranking in:

• LEADERBOARD.md
• reports/v0.4.0/clean_leaderboard.md
• reports/v0.4.0/noisy_leaderboard.md
• reports/v0.4.0/clean_noisy_paired_leaderboard.md

Repository Map

reports/reference_generation/v0.4.2/
  README.md
  index.html
  contact_sheet.png
  manifest.jsonl
  summary.json
  images_clean/
  metadata/
  thumbs/

reports/v0.4.0/
  clean_leaderboard.*
  noisy_leaderboard.*
  clean_noisy_paired_leaderboard.*

docs/releases/
  v0.4.0.md
  v0.4.1.md
  v0.4.2.md

assets/
  jomb_v042_receipt_wall.png

Quick Start

List records from the frozen dataset:

python examples/load_v020_manifest.py --data-root "data\v0.2.0" --limit 5 --show-paths

Evaluate your own prediction JSON files:

python examples/evaluate_v020_baseline.py --data-root "data\v0.2.0" --prediction-dir ".\model_outputs\my-model"

Open the v0.4.2 receipt structure gallery:

reports/reference_generation/v0.4.2/index.html

Data Policy

The public benchmark data is synthetic. Development references may be used to study receipt layout conventions, but public files must not include real customer receipt images, real personal information, copied brand logos, or local absolute paths.

License

See LICENSE.md in the public payload when mirrored. This Hugging Face dataset repository contains release artifacts, leaderboards, and sanitized synthetic receipt data for the Japan OCR Mini Benchmark project.