Multilingual training and evaluation data for AI teams
Native speakers label text, record and transcribe speech, rate model answers and red-team your model in the languages your users speak. You start with a paid pilot on your own data and your own guidelines, and scale once the agreement numbers are where you need them.
Three annotators label every pilot item.Disagreements traced to two unclear intents.Guideline updated with examples. Agreement rises.Weekly production batches, with agreement checked on each one.
Example of a pilot. Figures are illustrative.
- Data types
- Text, documents, speech, images, video and model outputs
- Work
- Annotation, collection, evaluation, RLHF and red teaming
- People
- Native speakers, matched to language, dialect and subject
- Start
- A paid pilot batch on your data and guidelines
Why AI teams come to a language company
Most annotation vendors are strongest in English. When your model has to work in Arabic dialects, Hindi with English mixed in, or formal and informal Japanese, you need annotators who are native in those languages and a team that knows how to write guidelines across them. That's what a translation company does every day.
We run your project like a translation project: one project manager, written guidelines, a second person checking the work, and numbers you can see on every batch.
What we do for AI teams
Each service has its own page with formats, steps and examples.
How a project starts
You see real output and real agreement numbers before you commit to volume.
Scoping call
Your task, languages, volumes, tool and how you'll judge quality.
Guidelines
We review your guidelines, flag gaps for each language and agree on edge cases.
Pilot batch
A few hundred items per language, labeled by several annotators, with an agreement report.
Agreement report QACalibration
Guidelines updated from the pilot, and annotators retrained on the changes.
Production
Regular batches in your tool or format, with quality checks on each one.
What you get on every batch
So you can decide whether the data is good enough to train on.
- The data in your format. JSONL, CSV or exported from your labeling tool.
- Agreement numbers. Agreement between annotators per language, with the method stated.
- A list of hard cases. Items annotators disagreed on, with notes for your team.
- Guideline questions. Gaps we found, so the next batch is cleaner.
Questions from AI teams
Which languages can you cover?
Major European and Asian languages, Arabic and its main dialects, Indic languages, and less common languages on request. Tell us the languages and dialects in the scoping call, and we'll confirm annotator availability for each.
Can you work in our labeling tool?
Yes. Our annotators can work in your own tool or a hosted platform such as Label Studio, or we can deliver files in the format your pipeline reads.
How does the pilot batch work?
We label a few hundred items per language from your own data, using your guidelines, with several annotators per item. You get the labeled data and an agreement report, and we agree any guideline changes before production.
How do you protect our data?
We sign your NDA, annotators work under confidentiality agreements, and access is limited to the people on your project.
Start with a pilot batch
Tell us the task, languages and volumes, or book a call. We reply with a pilot plan and a price.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time