| |
| """ |
| Task: ad-policy-compliance |
| Difficulty: Medium-Hard |
| Content types: ad_copy only |
| Steps per episode: 10 |
| Description: |
| Agent reviews ad copy and must: |
| 1. Identify violation labels |
| 2. Cite the specific AD_POLICY_* rule IDs that are violated |
| 3. Choose correct enforcement action (remove vs restrict vs approve) |
| Partial credit for correct label even without correct policy citation. |
| Full credit requires both correct label AND correct policy_citations. |
| """ |
| from __future__ import annotations |
| from models import ContentItem, ModerationObservation |
| from server.dataset import get_ad_copies, get_ad_policies |
|
|
| TASK_NAME = "ad-policy-compliance" |
| MAX_STEPS = 10 |
| INSTRUCTIONS = ( |
| "You are a Meta Ads policy reviewer. For each ad, you must: " |
| "1. Identify ALL violation categories present. " |
| "2. Cite the specific policy rule IDs (e.g. AD_POLICY_HEALTH_CLAIMS_001) that are violated. " |
| "3. Choose the correct action: approve (compliant), remove (clear violation), restrict (borderline). " |
| "The full policy document is provided in policy_excerpt. " |
| "Respond with a JSON ModerationDecision including policy_citations." |
| ) |
|
|
|
|
| def build_episode(seed: int = 42) -> list[ContentItem]: |
| ads = get_ad_copies(seed)[:MAX_STEPS] |
| return [ContentItem(**{k: v for k, v in ad.items() |
| if k not in ("ground_truth_labels", "ground_truth_action", |
| "difficulty", "violated_policies", "ad_targeting")}) |
| for ad in ads] |
|
|
|
|
| def build_observation(step: int, item: ContentItem) -> ModerationObservation: |
| policies = get_ad_policies() |
| full_policy_text = "\n".join( |
| f"[{p['id']}] {p['name']}: {p['description']}" |
| for p in policies["policies"] |
| ) |
| return ModerationObservation( |
| step=step, |
| content_item=item, |
| policy_excerpt=full_policy_text, |
| task_name=TASK_NAME, |
| instructions=INSTRUCTIONS, |
| ) |
|
|
