Custom handwriting datasets
Public handwriting datasets are where most handwriting recognition work begins. When a model needs writers from particular countries, particular words or phrases, a particular capture setup, or a licence that covers commercial training, the answer is a commissioned collection. GoKuna collects handwriting samples to your brief, written by adult contributors and reviewed before they count.
When a public handwriting dataset is not enough
Public datasets are fixed. Their writers, their text and their scanning setup were chosen for their original purpose, and each comes with its own licence terms, so check whether those cover commercial training.
A commissioned dataset lets you choose: the phrases written, how many writers, which countries they are in, and how each page is photographed. If your product will read handwriting from phone photos, it can learn from phone photos.
How GoKuna's handwriting task works
Contributors receive short phrases in the brief, write them in their own hand, and photograph the page flat in good light. The phrases contain nothing personal, so samples carry the handwriting, not details about the writer.
Contributors see the reward before they reserve the task; the current handwriting task pays $8.00 per accepted sample and takes about 5 minutes. Each sample is reviewed against the brief before it is accepted.
What you can specify
The text to be written, word for word. How it should be laid out on the page. How the photograph should be taken: flat, framed, lit. The accepted image formats and maximum file size. The number of samples. The countries writers must be in.
Each requirement becomes an instruction contributors read before they commit, and a rule reviewers check every sample against.
What makes handwriting data useful
A recognition model generalises from variety: many writers, each with their own letter shapes, slant, spacing and pressure, on different paper, in different light. A few hundred neat pages from a handful of writers teach it much less than the same number of pages from many ordinary hands.
Consistency matters as much as variety. When every sample follows the same phrases and the same photography rules, the only thing that changes from one sample to the next is the handwriting, which is exactly what the model is meant to learn.
Public and commissioned data together
Many teams use both. A public dataset gives a model its first exposure to handwriting; a commissioned collection then fills the gaps that matter for the product, such as the words it will actually read, the regions its users are in, or photos taken on phones rather than flatbed scans.
Because a commissioned brief fixes the text, you can also build evaluation sets where you know exactly what each page should say, and compare how the model reads the same phrases across many writers.
Review against your requirements
A person reviews every sample: are both phrases present and exactly as given, fully in frame, sharp and readable, with nothing extra on the page? Samples that miss are rejected with a reason, and the contributor can submit again while the task is open.
Contributors are paid only for accepted samples, which keeps attention on getting each page right.
Consent and licensing
Contributors are adults who confirm they are 18 or older and agree to only submit media they are authorized to share. Each task records the consent terms that applied to it.
Contributors keep ownership of their samples and grant a licence for training and evaluating machine learning models, including on behalf of GoKuna's customers. Accepted samples are kept as part of the delivered dataset. Your organisation is not named to contributors.
How to commission a handwriting dataset
Send a brief from the homepage form with your email and a description: the text, the number of samples, the regions, how you will use the data, and any budget or timing. GoKuna replies by email with a quote, and scope and price are agreed before collection starts.
If you are not sure how many writers or samples you need, say what the model is for and settle the numbers by email.
Can I choose the words that are written?
Yes. The phrases come from the brief, and contributors copy them exactly.
Can I get handwriting from specific countries?
Yes. A task can be limited to contributors in specific countries.
Are the samples scans or photos?
Phone photographs of the page, taken flat in good light, as the brief describes.
What are contributors paid for a handwriting sample?
The current handwriting task pays contributors $8.00 per accepted sample, shown before they reserve and paid through Stripe once the sample is accepted.
How is each sample checked?
A person reviews every sample against the brief before it is accepted. Samples that miss are rejected with a reason, and the contributor can submit again while the task is open.
Do samples contain personal information?
Contributors write the given phrases, which contain nothing personal, and are told to keep private information out of frame.