Multilingual data from professional translators.
Machine translation sounds fluent long before it is accurate. We commission professional translators to write and judge the translations multilingual models learn from.
Translation, with proof.
Machine translation and multilingual models fail in ways only a professional translator reliably catches: a dropped negation, a mistranslated term of art, a gender or formality choice that changes the meaning, a literal rendering of an idiom, terminology that drifts across a document. HUMXN commissions credential-checked translators to produce human translations, post-edit machine output, annotate errors by type and severity, and judge competing translations, in the language pairs and domains you specify.
Source texts come from consenting creators or are written new, with consent bound to each item. Documents are scanned for personal information and masked, fingerprinted at intake and checked for duplicates, so the same sentence does not appear in both training and evaluation data. A second translator reviews every item, label provenance records who made and confirmed each judgment, and deliveries carry a signed receipt.
For translators, this is paid, remote work in your own language pairs and specialist domains. You translate, post-edit, annotate errors and review a colleague's work, at a rate shown before you accept each job. Legal, medical and technical specialists are especially in demand. Your identity is not published, and approved work is paid automatically through Stripe after identity and tax verification.
Built to your specification.
Every engagement starts from a written spec and a pilot batch. These are the most common requests we source for translation.
Human reference translations
Professional translations of consented or newly written source text, scoped by language pair, domain and register.
Post-editing data
Machine output corrected by translators, with both versions kept so models learn from the edits.
Error annotation
Translation errors marked by span, category and severity using MQM-style typologies or your own scheme.
Quality and preference judgments
Translators rank competing translations for accuracy, fluency and terminology, with written rationales.
Terminology and glossaries
Domain glossaries and term-consistency checks for legal, medical, financial and technical content.
Localization evaluation
Checks for cultural fit, formality, date and number formats and locale conventions, beyond word-level accuracy.
01Fluent but wrong
Automatic metrics reward fluency. Professional judgment catches dropped negations, wrong terms and meaning shifts.
02Test set leakage
Parallel text is widely shared. Fingerprinting and duplicate checks keep evaluation sentences out of training data.
03Rights to source text
Translated works carry rights for author and translator. Consent bound to each item keeps the chain clear.
Paid work in your field.
Remote and flexible, with the rate stated before you accept. Every job is reviewed by a second expert, and approved work is paid automatically.
- A degree in translation, languages or a related field, or equivalent professional experience
- Several years of professional translation in your language pairs
- Certification from a recognized translators' association, or sworn translator status
- Native command of your target language
- Specialist domains such as legal, medical, financial or technical translation
Translate and post-edit
Produce reference translations, or correct machine output to a publishable standard.
Annotate translation errors
Mark error spans by category and severity following a written typology.
Judge competing translations
Rank alternatives for accuracy, fluency and terminology, and explain your choice.
Review another translator's work
Act as second reviewer on a colleague's translations and annotations before delivery.
Translation: asked often.
Which language pairs do you support?
Requests are scoped by language pair, direction, domain and register, then matched to translators verified in that pair and domain. Coverage depends on the request, so we confirm it during scoping and start with a pilot batch so your team can check quality, the error typology and the format.
How do you keep evaluation data separate from training data?
Every source segment and translation is fingerprinted at intake and checked for duplicates and near-duplicates across the dataset. Evaluation sets can be held back, licensed exclusively and timestamped, so you can show they existed before a model was trained and were not reused.
Do I need to be a certified translator?
Certification helps but is not required for every job. Professional experience in your pairs, native command of the target language and domain knowledge also qualify you. We verify identity and credentials, may set a short, paid test translation, and only offer jobs that match your verified pairs.
How do I get paid?
Rates are set by field and task and are shown before you accept, so you know what a job pays up front. Once a second translator approves your work, it is credited to your balance and paid automatically through Stripe after your identity and tax details are verified.
How much time is required?
There is no minimum commitment. Opportunities are offered when they match your pairs and domains, each with a stated scope and deadline, and you choose which to accept. The work is remote and fits around existing client work, so you can take on more during quiet periods and decline when you are busy.
Can we get exclusive rights to translations?
Yes. Commissioned translations and annotations can be licensed exclusively, with scope, term and exclusivity agreed in writing before work begins. Deliveries come as JSONL, CSV, Parquet or Croissant with a signed receipt, and standing orders can add new pairs or domains over time.
Translation, done by experts.
Need expert-made data in this field, or have the expertise to make it? Start here.