dataset card

README.md

@@ -75,7 +75,7 @@ target; it is eligible at drafting time (before the reply) but not in the pre-su
 All files of a release live under `<version>/` (this card: `v1.0/`); the git tag `v1.0` freezes the data. The
 **evidence supplement** (every run with per-query metrics and meta, all LLM-audit records including the extended
 audit and pools, and the fine-tuned bi-encoder weights) is under `v1.0/experiments/` and frozen by the tag
-`v1.0-evidence-r2` (the earlier `v1.0-evidence` lacks the audit of the random-negative models). Embedding caches are not included; the pipeline rebuilds them.
+`v1.0-evidence-r3` (earlier evidence tags lack the audit of the random-negative models and the post-hoc Qwen3 and dense question→question runs). Embedding caches are not included; the pipeline rebuilds them.
 
 - `corpus.jsonl`: `_id` (`<QID>-ASW`), `text` (masked, canonical), `raw_text` (header-stripped, unmasked),
   `date`, `institution`, `answers` (question ids), `header`, `source_url`.
@@ -111,6 +111,9 @@ controls (3 seeds each): random negatives 0.669 ± 0.012, random negatives withi
 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.
+Post-hoc zero-shot baseline: Qwen3-Embedding-8B (2025) reaches 0.714, with the same gain over BM25 (+0.17) on test
+questions answered after the model's release as before it. For CiteRet, dense question→question retrieval (BGE-M3)
+reaches R@10 0.744, on par with BM25 question→question (0.724).
 
 Splits are temporal (train 2014–2021, dev 2022–2023, test 2024–2026); 2015 questions are dated from their XML
 header.