mrshu commited on
Commit
4e39b10
·
verified ·
1 Parent(s): c84440c

dataset card

Browse files
Files changed (1) hide show
  1. README.md +10 -3
README.md CHANGED
@@ -65,6 +65,11 @@ All tasks have qrels files `qrels/<task>_<split>.tsv` (splits `train`, `dev`, `t
65
  - **Pre-submission variant** (`precedent`, `precedenthc`): same citation edges, eligible = answers published before
66
  the question date.
67
 
 
 
 
 
 
68
  ## Files
69
 
70
  All files of a release live under `<version>/` (this card: `v1.0/`); the git tag `v1.0` freezes them. Runs and
@@ -99,9 +104,11 @@ alone scores MRR@10 0.989) and a 0–180-day date window lifts BM25 from 0.547 t
99
  as-of filters of the citation tasks.
100
 
101
  **Baseline protocol caveat.** Fine-tuning on the training split with BM25 hard negatives mined across the whole
102
- archive (the pre-registered protocol) hurts (MRR@10 0.433 vs 0.490 zero-shot), consistent with a temporal
103
- negative-selection shortcut; date-matched negatives give 0.632, a post-hoc single-seed result to be confirmed on the
104
- next test period. Best pre-registered baseline: RRF(BM25, BGE-M3), 0.588.
 
 
105
 
106
  Splits are temporal (train 2014–2021, dev 2022–2023, test 2024–2026); 2015 questions are dated from their XML
107
  header.
 
65
  - **Pre-submission variant** (`precedent`, `precedenthc`): same citation edges, eligible = answers published before
66
  the question date.
67
 
68
+ **Example (test).** Question `P-9-2024-000684` (4 Mar 2024, asbestos screening proposal): its official reply
69
+ `P-9-2024-000684-ASW` is a referral stub ("…would refer the Honourable Member to its answer to written question
70
+ [REF]"), which is the AnswerRet target. The answer it cites, `E-9-2023-003699-ASW` (12 Mar 2024), is the CiteRet
71
+ target; it is eligible at drafting time (before the reply) but not in the pre-submission variant (after the question).
72
+
73
  ## Files
74
 
75
  All files of a release live under `<version>/` (this card: `v1.0/`); the git tag `v1.0` freezes them. Runs and
 
104
  as-of filters of the citation tasks.
105
 
106
  **Baseline protocol caveat.** Fine-tuning on the training split with BM25 hard negatives mined across the whole
107
+ archive (the pre-registered protocol) hurts (MRR@10 0.424 ± 0.008 over 3 seeds vs 0.490 zero-shot). Post-hoc
108
+ controls: random negatives 0.666, random negatives within ±365 days 0.708, date-matched BM25 negatives 0.637. The
109
+ mined negatives are mostly at least partially relevant answers the qrels miss (audited test queries: 18% "directly
110
+ addresses", 51% "partially"), plus a date signal. Avoid naive hard-negative mining here; these results are post hoc,
111
+ to be confirmed on the next test period. Best pre-registered baseline: RRF(BM25, BGE-M3), 0.588.
112
 
113
  Splits are temporal (train 2014–2021, dev 2022–2023, test 2024–2026); 2015 questions are dated from their XML
114
  header.