add the MTEB aggregate view and the order-insensitive ceiling
Browse files
sugarcrepe_symmetry_audit.json
CHANGED
|
@@ -84,5 +84,16 @@
|
|
| 84 |
]
|
| 85 |
}
|
| 86 |
},
|
| 87 |
-
"scope": "this is a property of the benchmark data alone. No model was run. It is not a defect: the swap splits are meant to probe word order. It is a caveat about how their scores read."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
}
|
|
|
|
| 84 |
]
|
| 85 |
}
|
| 86 |
},
|
| 87 |
+
"scope": "this is a property of the benchmark data alone. No model was run. It is not a defect: the swap splits are meant to probe word order. It is a caveat about how their scores read.",
|
| 88 |
+
"mteb_aggregate": {
|
| 89 |
+
"dataset": "mteb/SUGARCREPE_fmt",
|
| 90 |
+
"note": "MTEB merges all seven source subsets into one 7511-item test split, so the order-only items are a share of the whole benchmark.",
|
| 91 |
+
"n": 7511,
|
| 92 |
+
"bag_identical": 572,
|
| 93 |
+
"frac": 0.0762,
|
| 94 |
+
"replace_family_n": 3846,
|
| 95 |
+
"add_family_n": 2754,
|
| 96 |
+
"order_insensitive_ceiling": 0.9619,
|
| 97 |
+
"ceiling_note": "a model that ignores word order scores chance on the 572 order-only pairs, so its maximum text_acc is (7511-572)/7511 + 0.5*572/7511."
|
| 98 |
+
}
|
| 99 |
}
|