Read Any Language, Even in Bad Handwriting: An OCR Data Quotation, Teardown-Style

Every OCR project’s roadmap dies at the same milestone — “works on our test set, fails on real documents” — and the failure traces to training data that was clean, single-language, and printer-perfect while reality is multilingual, skewed, and photographed on a table. This piece takes a realistic RFQ response for labeled multilingual OCR data and tears it down line by line: which clauses are good, which are traps, and what a defensible quote looks like.

The RFQ That Started It

A fintech expanding across Southeast Asia needed OCR that reads identity documents, utility bills, and bank statements in five scripts, photographed by users under real-world conditions. Their request to a dataset vendor: 500,000 labeled document images, multilingual, with ground-truth transcriptions. The response arrived as a quotation — and the interesting part of this piece is what the quotation *should* contain, clause by clause, so you can tear down your own.

Clause 1: “500,000 labeled images”

The teardown: the number is the least informative part. What matters is the composition underneath it:

Good answer: a composition table with per-slice counts, and a note that the degradation mix targets field-captured rather than studio conditions. Trap: a single number and the word “diverse.”

On the shelf today, Thordata’s datamall carries multilingual OCR image lines with labels — priced around ¥4.5 per image on the published product tier (confirm current rates when scoping) — which for a 500K-image requirement frames the budget conversation in the low-to-mid six figures. That sounds like a lot until clause 2.

Clause 2: “Delivery in 4–6 weeks”

The teardown: calendar lines hide the real comparison. The in-house alternative to that quote — recruiting labelers across five scripts, building QA for ground-truth transcription, and collecting representative documents — is a two-quarter project with a hiring budget before a single image is labeled. The quotation’s honest rival isn’t “faster delivery”; it’s “you weren’t going to finish.” Six weeks for labeled, QA’d, composition-controlled data versus a quarter-plus of build risk is the actual arithmetic.

Clause 3: “Ground-truth accuracy 99%+”

The teardown: ask how it’s measured. Vendor-side QA on a sample? Double-annotation with adjudication? Inter-annotator agreement stats? The number is meaningless without methodology — and the follow-up that matters: *can you audit a random slice before purchase?* Sample-first evaluation is the category norm (Thordata’s flow lets you inspect sample fields before any commitment); a quotation that doesn’t offer it is a quotation that’s afraid of its own data.

The audit that works: pull 100 random images, transcribe them yourself with your best internal reviewer, diff against the vendor’s labels. Two hours, and it converts “99%+” from marketing into a measured number you can defend to your model team.

Clause 4: “License: internal training use”

The teardown: internal training covers the fine-tune. It does not automatically cover (a) model commercialization — if the OCR ships inside your product, does the license follow the model? — or (b) derived-data redistribution. For a fintech, clause 4’s precise wording is the clause their counsel will read three times. Good answer: license scope named per use — training, evaluation, commercial deployment — with the training-data question for deployed models addressed explicitly. Trap: “standard terms.”

Clause 5: “Refresh and expansion available on request”

The teardown: OCR deployments drift — new document formats, new languages, new phone cameras. The clause to demand is a *priced path*: extension slices at stated per-image rates, a cadence for new-language scoping, and delivery formats that match the first batch (JSON/CSV/Parquet, S3 delivery as standard). An unpriced “on request” is a future negotiation you’ve already lost.

The Reassembled Quote

What the defensible version looks like, summarized:

500,000 labeled document images
├─ composition: per-script × per-document-type × per-degradation counts
├─ labels: full transcription + region boxes, QA methodology documented
├─ audit: 100-image random slice pre-purchase, sample-first flow
├─ license: training + evaluation + deployed-model commercial use, named
├─ delivery: 4–6 weeks, JSON/CSV/Parquet, S3, schema documented
└─ extension: priced per-image refresh, new-language scoping path

None of these clauses are exotic. Their absence is what’s exotic — and expensive, discovered post-purchase.

FAQ

When is document-image collection better than a stock OCR corpus?
When your documents are specific — one country’s utility bills, one bank’s statements — the corpus covers the *general* reading skill and custom collection covers your formats. Most shipped systems are both: licensed breadth for the base model, targeted collection for the deployment domain.

How many images does an OCR fine-tune actually need?
Tens of thousands per script-family for a fine-tune on a strong base — the labeled-data question is quality-weighted, not volume-weighted. Composition (the clause-1 table) moves accuracy more than raw count; that’s why the teardown reads the number last.

Synthetic document generation instead?
Useful for structure and layout augmentation, insufficient alone: synthetic documents lack the *photographic* failure modes — glare, perspective, crumpled paper — that field images carry. The strongest training mixes licensed real images with synthetic augmentation, labeled as such.

What about the deployment’s feedback loop?
Underrated: your OCR in production produces rejected images daily, which are exactly the degradation cases the next training round needs. Pair the licensed corpus with a pipeline for routing production failures into an evaluation set — and watch the demand side too: which document types your users actually photograph, by market, shows up in search data before it shows up in your logs. The SERP monitoring solution — about $0.70 per 1,000 structured responses — reads that layer, and continuous SERP data crawling keeps the market picture current as your rollout adds countries.

What’s the one-sentence takeaway?
An OCR data quotation is readable in five clauses — composition, calendar, methodology, license, extension — and a vendor who answers all five in writing before you ask has already survived the teardown that matters; run query-side monitoring on your product’s search terms alongside, and you’ll know both what your model reads and what your market asks.