What you will do
You will read multilingual conversation prompts and label them for an AI-safety study. This is blind labelling: you assign your own judgments without seeing creator labels or other annotators’ answers.
MEAP
Multilingual Evaluation Annotation Platform
Welcome back
Sign in to continue annotating.
Forgot password?
Don't have an account? Sign up
MEAP
Reset your password
Forgot password
Enter your email and we'll send a reset link.
Remembered it? Sign in
MEAP
Set a new password
Create new password
Enter your new password below.
MEAP
Create your account
Join the platform
Complete your profile information first, then create your login credentials.
1. Profile information
2. Account credentials
Already have an account? Sign in
MEAP
Informed consent
Before you begin
Please read these points carefully. You can stop participating at any time.
Content warning
Prompts are designed to probe harmful AI behaviour and may include political, religious, or otherwise sensitive material. If you need to stop, close the session — there is no penalty.
Participation
Participation is voluntary and unpaid. You are not obliged to finish every item.
Credit & co-authorship
Stage 1 (Validation) completers are acknowledged. Stage 1 + Stage 2 (Evaluation) completers with good-quality work may be considered for co-authorship.
Account
Annotator ID:
Account information
Demographics
Change Password
My Annotations
Your submitted annotations across all prompts.
Queue progress
0%
👁
Blind labelling. Judge only what you see — language, turns, and optional English. Do not guess creator labels. Each prompt needs up to 2 independent annotations.
Your labels
Answer every required field. Use Skip only if you cannot annotate this item.
1
Annotator harm category * required
Your independent label. Creator categories stay hidden and are stored separately for later comparison.
2
Cultural specificity * required
Does understanding the harm require culturally specific knowledge (names, norms, local politics, religion), or is it universal?
3
Linguistic authenticity * required
Does the text sound natural in this language/dialect, or like awkward machine translation?
4
Cultural grounding * required
Is the conversation grounded in the target culture (setting, references, social context), not a generic English idea pasted in?
5
Escalation quality * required
Do the user turns build naturally toward a harmful ask, or does the jump feel forced / broken?
6
Severity * required
Rate potential real-world harm if an AI complied. Do not average — pick one level.
Low
Limited impact
Limited impact
Medium
Could mislead / offend / harm individuals
Could mislead / offend / harm individuals
High
Serious real-world harm
Serious real-world harm
7
Comments optional
Anything a reviewer should know — edge cases, uncertainty, dialect notes.
Benign twin check
This prompt includes a second, intended-harmless conversation. Compare dialect/register and confirm it is clearly harmless.
Same dialect, register, and tone? * required
Is the twin clearly harmless? * required
Twin comments optional
Complete required fields, then submit. Skip if you cannot annotate.
Create a prompt
Write a realistic multi-turn user conversation for AI-safety evaluation. An admin must approve it before it enters the annotation queue.
Provide user turns 1, 3, and 5 (assume the model replied in 2 and 4). English equivalents are required for every turn. Keep dialect natural.
Stored on the prompt only. Annotators never see this — they assign their own category later.
Example Prompt ()
Randomly selected from existing prompts in the dataset.
My Submitted Prompts
Admin Panel
Review prompts, manage users, and monitor annotation quality.
Starting prompts are pre-approved seed data. Only user-created submissions can be approved or rejected.
| ID | Segment | Language | Harm category | Turn 1 + English | Status | Annotations | Created | Actions |
|---|
No prompts match the current filters.
Registered Users
| Name | Role | Language | Annotator ID | Joined | Actions |
|---|
Completed Annotations
| Prompt ID | Annotator ID | Creator harm | Annotator harm | Severity | Ling. Auth | Cult. Spec | Timestamp | Actions |
|---|
Recycle Bin
Soft-deleted users, prompts, and annotations are kept here. Nothing is permanently removed. Restore anytime.
Analytics & Quality Control
Prompt detail
Create prompt
Create user
🎉
Queue complete!
There are no more prompts available for your language right now.
Connecting to server…