Datasets:
dataset card
Browse files
README.md
CHANGED
|
@@ -102,8 +102,9 @@ issues in v1.0").
|
|
| 102 |
|
| 103 |
**Permitted inputs.** Systems may use only the `text` fields. Ids, dates, `header`, `subject`, `raw_text`,
|
| 104 |
`source_url`, `answers` and the queries' `answer` field are forbidden: answer ids reuse the question id (that rule
|
| 105 |
-
alone scores MRR@10 0.989) and a 0–180-day date window lifts BM25 from 0.547 to 0.760.
|
| 106 |
-
as-of filters of the citation tasks
|
|
|
|
| 107 |
|
| 108 |
**Baseline protocol caveat.** Fine-tuning on the training split with BM25 hard negatives mined across the whole
|
| 109 |
archive (the pre-registered protocol) hurts (MRR@10 0.424 ± 0.008 over 3 seeds vs 0.490 zero-shot). Post-hoc
|
|
|
|
| 102 |
|
| 103 |
**Permitted inputs.** Systems may use only the `text` fields. Ids, dates, `header`, `subject`, `raw_text`,
|
| 104 |
`source_url`, `answers` and the queries' `answer` field are forbidden: answer ids reuse the question id (that rule
|
| 105 |
+
alone scores MRR@10 0.989) and a 0–180-day date window lifts BM25 from 0.547 to 0.760. At test time, dates may be used
|
| 106 |
+
only for the as-of filters of the citation tasks; training may use training-split metadata (two post-hoc recipes
|
| 107 |
+
select negatives by date).
|
| 108 |
|
| 109 |
**Baseline protocol caveat.** Fine-tuning on the training split with BM25 hard negatives mined across the whole
|
| 110 |
archive (the pre-registered protocol) hurts (MRR@10 0.424 ± 0.008 over 3 seeds vs 0.490 zero-shot). Post-hoc
|