Overview
This competition is a RAG task that can only be solved by using the fictional documents provided in the corpus. AI tools may be used freely, but solution models are limited to the 27B size.
Competition
Each test question requires a short free-form answer and a list of evidence sentence references from the released corpus.
Challenge removes several easy retrieval shortcuts and includes temporal, causal, bridge, comparison, and unanswerable cases.
The leaderboard is scored on the full 1,000-example organizer gold set. There is no separate public/private split.
Evaluation
The final score combines answer token F1 and evidence F1 using 0.7 * Answer F1 + 0.3 * Evidence F1. Submit predictions through the Submission page; only valid JSONL files are added to the leaderboard.
Participation
This competition is strictly individual. Use the assigned ID to log in, submit one prediction file at a time, and review your own submission history from the Submission page.