Trust and safety

Label harmful content against your own policy

Classify user content, write policy examples and build eval sets for your classifiers. Unrestricted models read and write the harmful examples a policy has to describe, within the Acceptable Use Policy.

Get API key Read the quickstart

Unrestricted means fewer refusals, not no rules. Read the Acceptable Use Policy

Labels, examples and eval items.

The policy and the label set are yours. The model applies them and explains its call.

  • /v1/chat/completions JSON mode

    A label with a reason

    Send your policy in the system prompt and one item per request. Get back the label, the clause and a short reason.

  • near-miss.jsonl Hard negatives

    Near-miss negatives

    Items that look like a violation but are allowed under your policy, to cut false positives.

  • policy/H2.md Policy examples

    Examples for every clause

    Violating, borderline and allowed examples, so reviewers and models apply a clause the same way.

How a policy team uses it.

You write the policy and own the decisions. The model does the first pass at volume.

JSON output guide

  1. You

    Write the policy as clauses

    Number each clause and define it, so labels can point back to it.

  2. Modellane

    Labels each item

    Label, clause, severity and reason, in the JSON shape you define.

  3. You

    Review the disagreements

    Cases where the model and your reviewers differ go to a human. Tighten the clause.

  4. Modellane

    Writes eval items

    Violating examples, near misses and held-out items for the clauses you changed.

What we do, and what we don't.

Modellane is an API you call from your own pipeline, not a moderation service.

Read the Acceptable Use Policy

What we do

  • Label content against the policy you provide
  • Write policy examples, near misses and eval items
  • Return labels as JSON in the fields you define

What we don't do

  • Provide a moderation service or a content policy of our own
  • Review content or make enforcement decisions for you
  • Store your prompts or the model's replies
  • Allow uses the Acceptable Use Policy prohibits, whatever the model answers

Which models fit: Models whose id contains -unrestricted read and write the harmful examples that typical assistants refuse to label or produce. Standard models keep their base model's behavior, behind the same API. See every model

FAQ

Trust and safety questions

Is this a moderation service?
No. It is an API. You send content and your policy, and you decide what to do with the label.
Can I use my own categories?
Yes. Define the labels and clauses in the system prompt; the model returns whatever fields you ask for.
Is the content I send stored?
No. Request content is processed in memory and discarded when the response ends.
Can labels go straight to enforcement?
Treat them as a first pass. Keep human review for appeals and for cases where your policy is unclear.

Start with the cases your reviewers argue about.

Create a key, paste your policy and label a batch.

Get API key Read the quickstart