This folder is a compact, human-inspected evidence subset copied from the official TEMPO and RECOR releases. It demonstrates that both proposed tracks have operational data formats, queries/conversations, complete positive evidence sets, and machine-readable qrels.
- This is pilot/trial material, not a new benchmark split and not the hidden SemEval-2027 test set.
- The selection is deliberately diverse and easy to inspect; it is not claimed to be statistically representative or domain-balanced.
- Every referenced positive document is included. No synthetic negatives, relevance labels, queries, answers, or conversation turns were added.
- The included document collection is only an evidence subset. Retrieval baselines and official scoring must use the full track corpus, not rank only these positive documents.
- All records are in English.
examples.jsonl: 5 complete temporal-retrieval examples.steps.jsonl: the matching official TEMPO step records.documents.jsonl: all 26 distinct gold passages referenced by the examples.qrels.txt: 26 positive query-document judgments in TREC format.
benchmark.jsonl: 4 complete conversational records containing 13 turns.documents.jsonl: all 15 distinct passages referenced by those turns.qrels.txt: 32 positive turn-document judgments in TREC format. Query IDs use<conversation_id>#turn<turn_id>.
manifest.json: counts, selected IDs, source URLs, and pinned revisions.