重构eval代码

2025-05-28 15:43:50 +08:00
parent 9f5318c23d
commit 9abd8fc1c5
39 changed files with 2468 additions and 166 deletions
--- a/results/20250528_1535/summary.json
+++ b/results/20250528_1535/summary.json
@@ -0,0 +1,60 @@
+{
+  "timestamp": "2025-05-28T15:36:05.540751",
+  "models_count": 2,
+  "models": {
+    "qwen-max-2025-01-25": {
+      "metrics": {
+        "accuracy": 1.0,
+        "precision_micro": 1.0,
+        "recall_micro": 1.0,
+        "f1_micro": 1.0,
+        "precision_macro": 1.0,
+        "recall_macro": 1.0,
+        "f1_macro": 1.0
+      },
+      "data_count": 10
+    },
+    "gpt-4o": {
+      "metrics": {
+        "accuracy": 1.0,
+        "precision_micro": 1.0,
+        "recall_micro": 1.0,
+        "f1_micro": 1.0,
+        "precision_macro": 1.0,
+        "recall_macro": 1.0,
+        "f1_macro": 1.0
+      },
+      "data_count": 10
+    }
+  },
+  "comparison": {
+    "accuracy": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    },
+    "precision_micro": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    },
+    "recall_micro": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    },
+    "f1_micro": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    },
+    "precision_macro": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    },
+    "recall_macro": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    },
+    "f1_macro": {
+      "qwen-max-2025-01-25": 1.0,
+      "gpt-4o": 1.0
+    }
+  }
+}