| |
| """ |
| Task: single-label-classify |
| Difficulty: Easy |
| Content types: text_post, image_description |
| Steps per episode: 10 |
| Description: |
| Agent receives one content item per step. |
| Must return exactly ONE ViolationCategory label. |
| Action must be one of: approve (if CLEAN) or remove/restrict. |
| """ |
| from __future__ import annotations |
| from models import ( |
| ContentItem, ModerationObservation, ContentType |
| ) |
| from server.dataset import get_posts, get_image_descriptions, get_policy_excerpt, get_community_standards |
|
|
| TASK_NAME = "single-label-classify" |
| MAX_STEPS = 10 |
| INSTRUCTIONS = ( |
| "You are a content moderator. For each item, assign EXACTLY ONE violation category " |
| "from the list, or mark it as CLEAN if no violation exists. " |
| "Then choose the appropriate enforcement action: approve (clean), remove (clear violation), " |
| "restrict (borderline), or escalate (unclear). " |
| "Respond with a JSON ModerationDecision." |
| ) |
|
|
|
|
| def build_episode(seed: int = 42) -> list[ContentItem]: |
| """Return a shuffled mix of posts and image descriptions for one episode.""" |
| posts = get_posts(seed)[:7] |
| images = get_image_descriptions(seed)[:3] |
| items = posts + images |
| return [ContentItem(**{k: v for k, v in item.items() |
| if k not in ("ground_truth_labels", "ground_truth_action", "difficulty")}) |
| for item in items] |
|
|
|
|
| def build_observation(step: int, item: ContentItem) -> ModerationObservation: |
| policies = get_community_standards() |
| return ModerationObservation( |
| step=step, |
| content_item=item, |
| policy_excerpt=get_policy_excerpt(item.content_type.value, policies), |
| task_name=TASK_NAME, |
| instructions=INSTRUCTIONS, |
| ) |
|
|
