Social platforms & marketplaces

Posts, listings and comments checked against your policy, live.

Policy labels on posts, listings, comments and live streams, with people reviewing classifier decisions in real time.

From upload to decision

Every part of the workflow where we can take work off your team.

Show
NSFW and nudity
Violence and gore
Hate and harassment

Policy labels

Classifier decisions checked
Live stream review
Appeals review

Live review

Harmful output flags
Red teaming
Refusal quality

Model safety

Prohibited listings
False claims and scams
Borderline case logs

Marketplaces

We work with images, video, live streams, text and comments.

What we label

Every class is written into a guide with your team before the first batch. Files come back as COCO JSON, YOLO, Pascal VOC, CSV, JSONL or your own schema.

Images and video

  • Policy categories

    Class and severity

    Nudity, violence, weapons, self harm

  • Regions

    Bounding box and blur mask

    Face, body, logo, text in image

Text and comments

  • Abuse

    Class and span

    Hate, harassment, spam, threat

Model output

  • Safety

    Rubric

    Harmful, biased, policy breaking, over refusal

How we test the model

The tests that matter for social platforms & marketplaces models, and what you get back from each.

Classifier precision

Removals and approvals sampled and rechecked by people.

You get backFalse removals and misses by policy

Red team findings

Adversarial prompts written to break your policy.

You get backFindings by severity, with examples

Reviewer agreement

Borderline content labelled twice, to find unclear policy lines.

You get backAgreement by policy category

Where social platforms & marketplaces teams start

Problems teams bring us, and the first piece of work we would scope.

Your classifier removes too much, or too little, and users notice both.

What we doPeople check classifier decisions before they reach users.Live review

Your policy has grey areas the model cannot learn from examples alone.

What we doBorderline cases labelled twice and logged, so the policy gets sharper.Annotate

Your model has to refuse harmful requests without refusing everything.

What we doRed teaming and refusal review against your policy.Evaluate & align

Related work

Client and volumes [REAL DATA: Client, volume and response time]

Bring us the social platforms & marketplaces data your model gets wrong.

Send a small batch and get it back done, free, usually within 24 to 48 hours. You judge the work before you pay for anything.