Based in Hong Kong · Serving AI teams worldwide
Asian AI data, with proof.
MXDATA produces, verifies and licenses rights-cleared datasets for AI, starting with Cantonese, Traditional Chinese and Hong Kong expert knowledge. Every dataset comes with its consent records and measured results.
Buying AI data is still a leap of faith
AI teams face three problems before they ever train on a dataset.
Rights you can't verify
Many "data owners" never secured the right to license their content for AI training. Buyers inherit that risk.
Languages models still miss
Cantonese, code-switched speech and Hong Kong domain knowledge are scarce in public data and hard to collect from outside the city.
Value you can't see before buying
A spec sheet doesn't tell you whether a dataset will improve your model on your task, or what it might degrade.
What we do
We produce our own data, carefully select a small number of partner datasets, and back each one with evidence.
Custom data production
Speech, text and expert data, collected from a vetted contributor network under explicit, documented consent and quality-checked at every step.
Datasets ready to license
Our own datasets, plus partner datasets that pass our rights review. Non-exclusive, custom and exclusive licences available.
Verification before you buy
Controlled experiments that show how a dataset changes a model on a defined task, with baselines, equal-budget controls and reported variance.
How we measure a dataset's effect
We don't hand out a single "quality score". We show what a dataset did, under which conditions, and at what cost.
Fix the setup
Every result is tied to a dataset version, model version, training recipe, test-set version and evaluation method.
Compare against controls
We compare against the base model and against public data trained on the same budget, so a gain from extra training isn't mistaken for a gain from better data.
Hold out properly
Test sets are split by speaker or by document, kept apart from training data, and never chosen by the data supplier.
Report everything
Metrics, variance, compute cost, failure cases and capabilities that got worse. If the difference is too small to call, we say so.
Results describe measured effects under the stated conditions. They are not a guarantee of gains for other models, tasks or data volumes.
MXDATA Certified
Every dataset we license ships with a documentation pack, so your legal and data teams can sign off quickly.
- Rights chainWho created the data and how rights reached us
- Consent recordsSigned, versioned contributor consent
- Personal data handlingDe-identification aligned with Hong Kong's PDPO
- Permitted useClear scope for training, evaluation and deployment
- Provenance logVersions, processing steps and QA reports
First datasets
We start where Hong Kong has an edge no one else can easily copy.
Hong Kong Cantonese–English code-switching speech
Natural two-speaker conversations in the way Hong Kong actually talks, for example "我 send 個 email 俾你 confirm 下".
- Everyday, customer-service, banking, healthcare and workplace scenarios
- 48 kHz, 16-bit, one channel per speaker
- Cantonese characters plus English, with timestamps and code-switch tags
- Balanced adult speakers; speaker-disjoint test split
Hong Kong expert evaluation sets
Questions, reference answers and grading rubrics written and reviewed by Hong Kong professionals.
- Finance and compliance, common-law legal reasoning, healthcare
- Traditional Chinese, Cantonese and English
- For model evaluation, preference data and red-teaming
Why Hong Kong
A trusted base for producing, certifying and licensing Asian data to the world.
Trilingual
Cantonese, Mandarin and English, in both Traditional and Simplified Chinese.
Expert depth
One of Asia's densest pools of finance, legal and medical professionals.
Common law
Internationally recognised contracts and IP framework.
Global settlement
Straightforward international contracting and payments.
Gateway to Asia
A hub for partners across Southeast and East Asia.
Talk to us
Looking for Asian-language data, want a sample of our first dataset, or have data or expertise to contribute? Tell us a little about your needs.