We mark up your images and video, saying what each thing is and where, so AI can learn to recognise it.
Explore serviceData & annotation
Human judgement that shapes how your AI behaves.
Chatbots and writing tools learn their tone, judgement and limits from examples that people write and rank. We provide trained reviewers and subject specialists who write, compare and critique an AI's answers against clear rules. Each task is tried on a small scale first and reviewed, so what your AI learns from stays consistent.
Want an AI agent tested end to end, through the channels your customers use?
See AI agent testingWhat we offer
Data Annotation for Generative AI services.
- 01
Preference ranking (RLHF)
Side-by-side comparisons and ratings of model responses, with written rationales where your process needs them.
- 02
Supervised fine-tuning data
Prompts and high-quality responses written to your style guide, covering the tasks and domains you care about.
- 03
Red-teaming & safety labelling
Adversarial prompts and safety labels that probe for harmful, biased or policy-breaking behaviour.
- 04
Evaluation datasets and rubrics
Test sets and scoring guides that measure model quality on the tasks that matter to your product.
- 05
Multimodal pairs
Image-caption pairs, visual question answering and other combined image and text data for multimodal models.
- 06
Domain-expert review
Specialist reviewers brought in for technical, legal, financial or other domain content where general reviewers are not enough.
How we work
From guidelines to delivered data.
AI takes the first pass, trained people review it, and quality is checked before delivery.
- 01
Scope
We agree label definitions, edge cases, formats and quality targets with you.
- 02
Pilot
A small batch tests the guidelines before any volume work starts.
- 03
Annotate
Trained annotators label to the agreed guidelines, with tooling to match.
- 04
Review
Multi-stage quality review against the specification, with feedback loops.
- 05
Deliver
Data in your format, with notes on what was checked, then we iterate.
Where it helps
Typical projects.
The kinds of work this service usually supports.
Aligning assistants
Collecting preference data that teaches an assistant to be helpful, accurate and appropriately cautious.
Domain fine-tuning
Building instruction datasets that adapt a base model to a specific industry, product or writing style.
Pre-launch safety testing
Red-teaming a model or feature before release and turning the findings into safety training data.
Ongoing model evaluation
Maintaining evaluation sets that let you compare model versions and prompts on consistent criteria.
Details
Output formats
- JSONL
- Chat message JSONL
- Pairwise preference files
- Rubric scores (CSV)
- Parquet
- Hugging Face datasets
- Image-caption manifests
- Custom schemas
Questions
Who does the writing and ranking?
Trained reviewers who work to your guidelines, with subject specialists brought in for technical or specialist content. Reviewers are calibrated on sample tasks before production so their judgements line up.
How do you keep ratings consistent?
We use written rubrics, calibration rounds and overlapping assignments to spot disagreement early. Disagreements are resolved and the guidelines updated, so the same question gets the same answer.
Can you handle sensitive or harmful content?
Yes, with care. Red-teaming and safety work is done by briefed reviewers, under NDA if needed, with content boundaries agreed before the work begins.
More data & annotation services
All services
Match footage turned into data: who was where, what happened and when, for coaches, scouts and media teams.
Explore service
Software that watches video for you and flags what matters, from recorded footage to live alerts.
Explore service
Idea → Architecture → Prototype → Production → Scale
Talk to us about Data Annotation for Generative AI.
Tell us what you're trying to build, test, automate or improve. We'll tell you honestly where we can make a real difference, and where we can't.