Spaces:
Sleeping
Sleeping
| from typing import Any, Dict | |
| from server.tasks.base_task import BaseTask | |
| class TriageTask(BaseTask): | |
| def grade(self, final_state: Dict[str, Any], org_config: Dict[str, Any], scenario: Dict[str, Any]) -> float: | |
| score = 0.0 | |
| expected = scenario["expected"] | |
| tickets = final_state["ticketing"]["tickets"] | |
| new_tickets = [t for t in tickets if t.get("created_this_episode")] | |
| if not new_tickets: | |
| return 0.0 | |
| score -= 0.10 * max(0, len(new_tickets) - 1) # duplicate penalty | |
| t = new_tickets[0] | |
| score += 0.25 # ticket exists | |
| if t.get("label") == expected.get("label"): | |
| score += 0.20 | |
| if t.get("priority") == expected.get("priority"): | |
| score += 0.20 | |
| if t.get("assigned_team") == expected.get("team"): | |
| score += 0.20 | |
| msgs = final_state["chat"]["messages_posted_this_episode"] | |
| correct_msgs = [m for m in msgs if m.get("channel") == expected.get("channel")] | |
| wrong_msgs = [m for m in msgs if m.get("channel") != expected.get("channel")] | |
| if correct_msgs: | |
| score += 0.10 | |
| if len(correct_msgs[0].get("text", "")) >= 20: | |
| score += 0.05 | |
| score -= 0.05 * len(wrong_msgs) | |
| return max(0.0, min(1.0, score)) | |