Create results/summary.json

results/summary.json

@@ -0,0 +1,62 @@
+{
+    "Classification": {
+        "amazon_counterfactual_classification": {
+            "macro_f1": 0.8035391985235245
+        },
+        "amazon_review_classification": {
+            "macro_f1": 0.5723768855928715
+        },
+        "massive_intent_classification": {
+            "macro_f1": 0.8139892125095518
+        },
+        "massive_scenario_classification": {
+            "macro_f1": 0.8973284738295028
+        }
+    },
+    "Reranking": {
+        "esci": {
+            "ndcg@10": 0.9321196064872125
+        }
+    },
+    "Retrieval": {
+        "jagovfaqs_22k": {
+            "ndcg@10": 0.781068526946658
+        },
+        "jaqket": {
+            "ndcg@10": 0.6555161728617205
+        },
+        "mrtydi": {
+            "ndcg@10": 0.46581796830150424
+        },
+        "nlp_journal_abs_intro": {
+            "ndcg@10": 0.8926829121538079
+        },
+        "nlp_journal_title_abs": {
+            "ndcg@10": 0.9777828877171447
+        },
+        "nlp_journal_title_intro": {
+            "ndcg@10": 0.8074433150584185
+        }
+    },
+    "STS": {
+        "jsick": {
+            "spearman": 0.8210090897926421
+        },
+        "jsts": {
+            "spearman": 0.8423674566905451
+        }
+    },
+    "Clustering": {
+        "livedoor_news": {
+            "v_measure_score": 0.5374210312008371
+        },
+        "mewsc16": {
+            "v_measure_score": 0.505362189310571
+        }
+    },
+    "PairClassification": {
+        "paws_x_ja": {
+            "binary_f1": 0.6226614895870103
+        }
+    }
+}