Datasets:
dataset card
Browse files
README.md
CHANGED
|
@@ -126,6 +126,13 @@ alone scores MRR@10 0.989) and a 0–180-day date window lifts BM25 from 0.547 t
|
|
| 126 |
only for the as-of filters of the citation tasks; training may use training-split metadata (two post-hoc recipes
|
| 127 |
select negatives by date).
|
| 128 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
**Safe default for evaluation: the text-only export** (tag `v1.0-textonly`, `v1.0/textonly/`): corpus and queries with
|
| 130 |
only a random opaque `_id` and the masked `text`; qrels on opaque ids; evaluator-only date files for the as-of rule in
|
| 131 |
`eval/`; the opaque→source mapping as a separate download (`v1.0/textonly-provenance/`). Every main-experiment run scores
|
|
|
|
| 126 |
only for the as-of filters of the citation tasks; training may use training-split metadata (two post-hoc recipes
|
| 127 |
select negatives by date).
|
| 128 |
|
| 129 |
+
**Citation-task allowance.** For `precedent*` and `cite*`, a system may also use the archive's question→answer links
|
| 130 |
+
of *eligible* answers (dated before the query's cutoff) and those earlier questions' `text`; the query's own `answer`
|
| 131 |
+
link and anything after its cutoff stay forbidden (the as-of filter enforces the dates). Only the question→question
|
| 132 |
+
baselines use this; all other baselines rank answer texts only. The text-only export does not contain these links,
|
| 133 |
+
so question→question baselines cannot be rerun from it (use the v1.0 data, commit
|
| 134 |
+
`2176502d8ed76885c17fc85604be33d2cf429368`).
|
| 135 |
+
|
| 136 |
**Safe default for evaluation: the text-only export** (tag `v1.0-textonly`, `v1.0/textonly/`): corpus and queries with
|
| 137 |
only a random opaque `_id` and the masked `text`; qrels on opaque ids; evaluator-only date files for the as-of rule in
|
| 138 |
`eval/`; the opaque→source mapping as a separate download (`v1.0/textonly-provenance/`). Every main-experiment run scores
|