dataset card

README.md

@@ -105,10 +105,10 @@ as-of filters of the citation tasks.
 
 **Baseline protocol caveat.** Fine-tuning on the training split with BM25 hard negatives mined across the whole
 archive (the pre-registered protocol) hurts (MRR@10 0.424 ± 0.008 over 3 seeds vs 0.490 zero-shot). Post-hoc
-controls: random negatives 0.666, random negatives within ±365 days 0.708, date-matched BM25 negatives 0.637. The
-mined negatives are mostly at least partially relevant answers the qrels miss (audited test queries: 18% "directly
-addresses", 51% "partially"), plus a date signal. Avoid naive hard-negative mining here; these results are post hoc,
-to be confirmed on the next test period. Best pre-registered baseline: RRF(BM25, BGE-M3), 0.588.
+controls: random negatives 0.666, random negatives within ±365 days 0.708 (one seed each), date-matched BM25
+negatives 0.637 (3 seeds). Why is not established; plausibly many mined negatives are relevant answers the qrels miss
+(LLM-judged test-query proxy: 18% "directly addresses", 51% "partially"), plus a date signal. Avoid naive
+hard-negative mining here; these results are post hoc, to be confirmed on the next test period. Best pre-registered baseline: RRF(BM25, BGE-M3), 0.588.
 
 Splits are temporal (train 2014–2021, dev 2022–2023, test 2024–2026); 2015 questions are dated from their XML
 header.