LR2Bench

Running

App Files Files Community

UltraRonin commited on 20 days ago

Commit

c5ebc43

1 Parent(s): ce9edf3

add

Browse files

Files changed (8) hide show

Acrostic.json +10 -10
Crossword.json +10 -10
Cryptogram.json +8 -8
Drop_Quote.json +10 -10
Logic_Puzzle.json +13 -13
Overall.json +130 -0
Sudoku.json +10 -10
index.html +1 -0

Acrostic.json CHANGED Viewed

@@ -2,7 +2,7 @@
     {
         "Model": "Llama-3.1-8B-Instruct",
         "CR": "43.0",
-        "S-Acc": "5.5",
         "EM": "0.0",
         "PM-0.5": "0.0",
         "Tokens": "3712"
@@ -29,7 +29,7 @@
         "S-Acc": "7.9",
         "EM": "0.0",
         "PM-0.5": "0.0",
-        "Tokens": "4599"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
@@ -37,7 +37,7 @@
         "S-Acc": "5.5",
         "EM": "0.0",
         "PM-0.5": "0.0",
-        "Tokens": "4170"
     },
     {
         "Model": "Mistral-Large-Instruct-2411",
@@ -69,7 +69,7 @@
         "S-Acc": "39.3",
         "EM": "0.0",
         "PM-0.5": "18.0",
-        "Tokens": "4110"
     },
     {
         "Model": "QwQ-32B-Preview",
@@ -85,23 +85,23 @@
         "S-Acc": "62.2",
         "EM": "0.0",
         "PM-0.5": "83.0",
-        "Tokens": "10076"
     },
     {
-        "Model": "Gemini-2.0-exp",
         "CR": "98.0",
         "S-Acc": "48.0",
         "EM": "0.0",
         "PM-0.5": "48.0",
-        "Tokens": "4019"
     },
     {
-        "Model": "Gemini-2.0-thinking",
         "CR": "92.0",
         "S-Acc": "40.7",
         "EM": "0.0",
         "PM-0.5": "27.0",
-        "Tokens": "4256"
     },
     {
         "Model": "OpenAI-gpt-4o",
@@ -117,7 +117,7 @@
         "S-Acc": "34.7",
         "EM": "0.0",
         "PM-0.5": "12.0",
-        "Tokens": "10951"
     },
     {
         "Model": "OpenAI-o1-preview",

     {
         "Model": "Llama-3.1-8B-Instruct",
         "CR": "43.0",
+        "S-Acc": "5.6",
         "EM": "0.0",
         "PM-0.5": "0.0",
         "Tokens": "3712"
         "S-Acc": "7.9",
         "EM": "0.0",
         "PM-0.5": "0.0",
+        "Tokens": "4600"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
         "S-Acc": "5.5",
         "EM": "0.0",
         "PM-0.5": "0.0",
+        "Tokens": "4171"
     },
     {
         "Model": "Mistral-Large-Instruct-2411",
         "S-Acc": "39.3",
         "EM": "0.0",
         "PM-0.5": "18.0",
+        "Tokens": "4111"
     },
     {
         "Model": "QwQ-32B-Preview",
         "S-Acc": "62.2",
         "EM": "0.0",
         "PM-0.5": "83.0",
+        "Tokens": "10077"
     },
     {
+        "Model": "Gemini-2.0-flash",
         "CR": "98.0",
         "S-Acc": "48.0",
         "EM": "0.0",
         "PM-0.5": "48.0",
+        "Tokens": "4020"
     },
     {
+        "Model": "Gemini-2.0-flash-thinking",
         "CR": "92.0",
         "S-Acc": "40.7",
         "EM": "0.0",
         "PM-0.5": "27.0",
+        "Tokens": "4257"
     },
     {
         "Model": "OpenAI-gpt-4o",
         "S-Acc": "34.7",
         "EM": "0.0",
         "PM-0.5": "12.0",
+        "Tokens": "10952"
     },
     {
         "Model": "OpenAI-o1-preview",

Crossword.json CHANGED Viewed

@@ -5,7 +5,7 @@
         "S-Acc": "23.3",
         "EM": "0.0",
         "PM-0.5": "14.0",
-        "Tokens": "2887"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
@@ -13,7 +13,7 @@
         "S-Acc": "46.8",
         "EM": "0.0",
         "PM-0.5": "62.0",
-        "Tokens": "3071"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
@@ -21,7 +21,7 @@
         "S-Acc": "47.6",
         "EM": "0.0",
         "PM-0.5": "65.3",
-        "Tokens": "2612"
     },
     {
         "Model": "Mistral-7B-Instruct-v0.3",
@@ -37,7 +37,7 @@
         "S-Acc": "48.3",
         "EM": "0.0",
         "PM-0.5": "54.0",
-        "Tokens": "3134"
     },
     {
         "Model": "Mistral-Large-Instruct-2411",
@@ -69,7 +69,7 @@
         "S-Acc": "44.1",
         "EM": "0.0",
         "PM-0.5": "36.7",
-        "Tokens": "2734"
     },
     {
         "Model": "QwQ-32B-Preview",
@@ -77,7 +77,7 @@
         "S-Acc": "30.2",
         "EM": "0.0",
         "PM-0.5": "18.0",
-        "Tokens": "4816"
     },
     {
         "Model": "DeepSeek-R1",
@@ -85,10 +85,10 @@
         "S-Acc": "75.3",
         "EM": "16.7",
         "PM-0.5": "94.0",
-        "Tokens": "9809"
     },
     {
-        "Model": "Gemini-2.0-exp",
         "CR": "98.7",
         "S-Acc": "61.6",
         "EM": "0.0",
@@ -96,7 +96,7 @@
         "Tokens": "2555"
     },
     {
-        "Model": "Gemini-2.0-thinking",
         "CR": "94.7",
         "S-Acc": "57.7",
         "EM": "1.3",
@@ -106,7 +106,7 @@
     {
         "Model": "OpenAI-gpt-4o",
         "CR": "100.0",
-        "S-Acc": "63.0",
         "EM": "1.3",
         "PM-0.5": "86.7",
         "Tokens": "1726"

         "S-Acc": "23.3",
         "EM": "0.0",
         "PM-0.5": "14.0",
+        "Tokens": "2888"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
         "S-Acc": "46.8",
         "EM": "0.0",
         "PM-0.5": "62.0",
+        "Tokens": "3072"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
         "S-Acc": "47.6",
         "EM": "0.0",
         "PM-0.5": "65.3",
+        "Tokens": "2613"
     },
     {
         "Model": "Mistral-7B-Instruct-v0.3",
         "S-Acc": "48.3",
         "EM": "0.0",
         "PM-0.5": "54.0",
+        "Tokens": "3135"
     },
     {
         "Model": "Mistral-Large-Instruct-2411",
         "S-Acc": "44.1",
         "EM": "0.0",
         "PM-0.5": "36.7",
+        "Tokens": "2735"
     },
     {
         "Model": "QwQ-32B-Preview",
         "S-Acc": "30.2",
         "EM": "0.0",
         "PM-0.5": "18.0",
+        "Tokens": "4817"
     },
     {
         "Model": "DeepSeek-R1",
         "S-Acc": "75.3",
         "EM": "16.7",
         "PM-0.5": "94.0",
+        "Tokens": "9810"
     },
     {
+        "Model": "Gemini-2.0-flash",
         "CR": "98.7",
         "S-Acc": "61.6",
         "EM": "0.0",
         "Tokens": "2555"
     },
     {
+        "Model": "Gemini-2.0-flash-thinking",
         "CR": "94.7",
         "S-Acc": "57.7",
         "EM": "1.3",
     {
         "Model": "OpenAI-gpt-4o",
         "CR": "100.0",
+        "S-Acc": "66.0",
         "EM": "1.3",
         "PM-0.5": "86.7",
         "Tokens": "1726"

Cryptogram.json CHANGED Viewed

@@ -5,7 +5,7 @@
         "S-Acc": "2.3",
         "EM": "0.0",
         "PM-0.5": "0.0",
-        "Tokens": "2067"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
@@ -13,7 +13,7 @@
         "S-Acc": "6.9",
         "EM": "0.0",
         "PM-0.5": "1.0",
-        "Tokens": "1297"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
@@ -29,7 +29,7 @@
         "S-Acc": "4.3",
         "EM": "0.0",
         "PM-0.5": "0.0",
-        "Tokens": "1095"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
@@ -69,7 +69,7 @@
         "S-Acc": "11.8",
         "EM": "0.0",
         "PM-0.5": "0.0",
-        "Tokens": "1726"
     },
     {
         "Model": "QwQ-32B-Preview",
@@ -77,7 +77,7 @@
         "S-Acc": "3.6",
         "EM": "0.0",
         "PM-0.5": "0.0",
-        "Tokens": "6491"
     },
     {
         "Model": "DeepSeek-R1",
@@ -88,7 +88,7 @@
         "Tokens": "10344"
     },
     {
-        "Model": "Gemini-2.0-exp",
         "CR": "47.0",
         "S-Acc": "8.5",
         "EM": "0.0",
@@ -96,7 +96,7 @@
         "Tokens": "1585"
     },
     {
-        "Model": "Gemini-2.0-thinking",
         "CR": "68.0",
         "S-Acc": "11.2",
         "EM": "0.0",
@@ -109,7 +109,7 @@
         "S-Acc": "20.7",
         "EM": "0.0",
         "PM-0.5": "5.0",
-        "Tokens": "739"
     },
     {
         "Model": "OpenAI-o1-mini",

         "S-Acc": "2.3",
         "EM": "0.0",
         "PM-0.5": "0.0",
+        "Tokens": "2068"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
         "S-Acc": "6.9",
         "EM": "0.0",
         "PM-0.5": "1.0",
+        "Tokens": "1298"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
         "S-Acc": "4.3",
         "EM": "0.0",
         "PM-0.5": "0.0",
+        "Tokens": "1096"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
         "S-Acc": "11.8",
         "EM": "0.0",
         "PM-0.5": "0.0",
+        "Tokens": "1727"
     },
     {
         "Model": "QwQ-32B-Preview",
         "S-Acc": "3.6",
         "EM": "0.0",
         "PM-0.5": "0.0",
+        "Tokens": "6492"
     },
     {
         "Model": "DeepSeek-R1",
         "Tokens": "10344"
     },
     {
+        "Model": "Gemini-2.0-flash",
         "CR": "47.0",
         "S-Acc": "8.5",
         "EM": "0.0",
         "Tokens": "1585"
     },
     {
+        "Model": "Gemini-2.0-flash-thinking",
         "CR": "68.0",
         "S-Acc": "11.2",
         "EM": "0.0",
         "S-Acc": "20.7",
         "EM": "0.0",
         "PM-0.5": "5.0",
+        "Tokens": "740"
     },
     {
         "Model": "OpenAI-o1-mini",

Drop_Quote.json CHANGED Viewed

@@ -5,7 +5,7 @@
         "S-Acc": "11.2",
         "EM": "0.0",
         "PM-0.5": "1.0",
-        "Tokens": "2122"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
@@ -29,7 +29,7 @@
         "S-Acc": "6.6",
         "EM": "0.0",
         "PM-0.5": "1.0",
-        "Tokens": "2336"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
@@ -37,7 +37,7 @@
         "S-Acc": "26.9",
         "EM": "0.0",
         "PM-0.5": "6.0",
-        "Tokens": "1614"
     },
     {
         "Model": "Mistral-Large-Instruct-2411",
@@ -45,7 +45,7 @@
         "S-Acc": "24.7",
         "EM": "0.0",
         "PM-0.5": "9.0",
-        "Tokens": "1565"
     },
     {
         "Model": "Qwen2.5-7B-Instruct",
@@ -53,7 +53,7 @@
         "S-Acc": "21.9",
         "EM": "0.0",
         "PM-0.5": "4.0",
-        "Tokens": "1851"
     },
     {
         "Model": "Qwen2.5-32B-Instruct",
@@ -69,7 +69,7 @@
         "S-Acc": "30.9",
         "EM": "0.0",
         "PM-0.5": "13.0",
-        "Tokens": "1756"
     },
     {
         "Model": "QwQ-32B-Preview",
@@ -88,20 +88,20 @@
         "Tokens": "11422"
     },
     {
-        "Model": "Gemini-2.0-exp",
         "CR": "92.0",
-        "S-Acc": "34.2",
         "EM": "0.0",
         "PM-0.5": "17.0",
         "Tokens": "2717"
     },
     {
-        "Model": "Gemini-2.0-thinking",
         "CR": "96.0",
         "S-Acc": "34.4",
         "EM": "0.0",
         "PM-0.5": "23.0",
-        "Tokens": "3385"
     },
     {
         "Model": "OpenAI-gpt-4o",

         "S-Acc": "11.2",
         "EM": "0.0",
         "PM-0.5": "1.0",
+        "Tokens": "2123"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
         "S-Acc": "6.6",
         "EM": "0.0",
         "PM-0.5": "1.0",
+        "Tokens": "2337"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
         "S-Acc": "26.9",
         "EM": "0.0",
         "PM-0.5": "6.0",
+        "Tokens": "1615"
     },
     {
         "Model": "Mistral-Large-Instruct-2411",
         "S-Acc": "24.7",
         "EM": "0.0",
         "PM-0.5": "9.0",
+        "Tokens": "1566"
     },
     {
         "Model": "Qwen2.5-7B-Instruct",
         "S-Acc": "21.9",
         "EM": "0.0",
         "PM-0.5": "4.0",
+        "Tokens": "1852"
     },
     {
         "Model": "Qwen2.5-32B-Instruct",
         "S-Acc": "30.9",
         "EM": "0.0",
         "PM-0.5": "13.0",
+        "Tokens": "1757"
     },
     {
         "Model": "QwQ-32B-Preview",
         "Tokens": "11422"
     },
     {
+        "Model": "Gemini-2.0-flash",
         "CR": "92.0",
+        "S-Acc": "34.3",
         "EM": "0.0",
         "PM-0.5": "17.0",
         "Tokens": "2717"
     },
     {
+        "Model": "Gemini-2.0-flash-thinking",
         "CR": "96.0",
         "S-Acc": "34.4",
         "EM": "0.0",
         "PM-0.5": "23.0",
+        "Tokens": "3386"
     },
     {
         "Model": "OpenAI-gpt-4o",

Logic_Puzzle.json CHANGED Viewed

@@ -5,7 +5,7 @@
         "S-Acc": "16.0",
         "EM": "0.0",
         "PM-0.5": "8.0",
-        "Tokens": "1292"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
@@ -13,7 +13,7 @@
         "S-Acc": "22.8",
         "EM": "2.0",
         "PM-0.5": "18.0",
-        "Tokens": "1164"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
@@ -29,7 +29,7 @@
         "S-Acc": "19.1",
         "EM": "0.0",
         "PM-0.5": "4.5",
-        "Tokens": "1617"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
@@ -45,7 +45,7 @@
         "S-Acc": "38.3",
         "EM": "3.0",
         "PM-0.5": "30.5",
-        "Tokens": "1636"
     },
     {
         "Model": "Qwen2.5-7B-Instruct",
@@ -53,7 +53,7 @@
         "S-Acc": "25.8",
         "EM": "0.0",
         "PM-0.5": "8.5",
-        "Tokens": "1395"
     },
     {
         "Model": "Qwen2.5-32B-Instruct",
@@ -61,7 +61,7 @@
         "S-Acc": "32.2",
         "EM": "0.0",
         "PM-0.5": "22.5",
-        "Tokens": "1207"
     },
     {
         "Model": "Qwen2.5-72B-Instruct",
@@ -69,7 +69,7 @@
         "S-Acc": "34.0",
         "EM": "0.0",
         "PM-0.5": "23.0",
-        "Tokens": "1809"
     },
     {
         "Model": "QwQ-32B-Preview",
@@ -77,7 +77,7 @@
         "S-Acc": "46.3",
         "EM": "19.5",
         "PM-0.5": "48.0",
-        "Tokens": "9523"
     },
     {
         "Model": "DeepSeek-R1",
@@ -85,23 +85,23 @@
         "S-Acc": "69.4",
         "EM": "42.5",
         "PM-0.5": "68.0",
-        "Tokens": "9204"
     },
     {
-        "Model": "Gemini-2.0-exp",
         "CR": "58.0",
         "S-Acc": "24.2",
         "EM": "2.0",
         "PM-0.5": "20.0",
-        "Tokens": "2103"
     },
     {
-        "Model": "Gemini-2.0-thinking",
         "CR": "99.0",
         "S-Acc": "45.9",
         "EM": "8.0",
         "PM-0.5": "37.5",
-        "Tokens": "4037"
     },
     {
         "Model": "OpenAI-gpt-4o",

         "S-Acc": "16.0",
         "EM": "0.0",
         "PM-0.5": "8.0",
+        "Tokens": "1293"
     },
     {
         "Model": "Llama-3.1-70B-Instruct",
         "S-Acc": "22.8",
         "EM": "2.0",
         "PM-0.5": "18.0",
+        "Tokens": "1165"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
         "S-Acc": "19.1",
         "EM": "0.0",
         "PM-0.5": "4.5",
+        "Tokens": "1618"
     },
     {
         "Model": "Mistral-Small-Instruct-2409",
         "S-Acc": "38.3",
         "EM": "3.0",
         "PM-0.5": "30.5",
+        "Tokens": "1637"
     },
     {
         "Model": "Qwen2.5-7B-Instruct",
         "S-Acc": "25.8",
         "EM": "0.0",
         "PM-0.5": "8.5",
+        "Tokens": "1396"
     },
     {
         "Model": "Qwen2.5-32B-Instruct",
         "S-Acc": "32.2",
         "EM": "0.0",
         "PM-0.5": "22.5",
+        "Tokens": "1208"
     },
     {
         "Model": "Qwen2.5-72B-Instruct",
         "S-Acc": "34.0",
         "EM": "0.0",
         "PM-0.5": "23.0",
+        "Tokens": "1810"
     },
     {
         "Model": "QwQ-32B-Preview",
         "S-Acc": "46.3",
         "EM": "19.5",
         "PM-0.5": "48.0",
+        "Tokens": "9524"
     },
     {
         "Model": "DeepSeek-R1",
         "S-Acc": "69.4",
         "EM": "42.5",
         "PM-0.5": "68.0",
+        "Tokens": "9205"
     },
     {
+        "Model": "Gemini-2.0-flash",
         "CR": "58.0",
         "S-Acc": "24.2",
         "EM": "2.0",
         "PM-0.5": "20.0",
+        "Tokens": "2104"
     },
     {
+        "Model": "Gemini-2.0-flash-thinking",
         "CR": "99.0",
         "S-Acc": "45.9",
         "EM": "8.0",
         "PM-0.5": "37.5",
+        "Tokens": "4038"
     },
     {
         "Model": "OpenAI-gpt-4o",

Overall.json ADDED Viewed

	@@ -0,0 +1,130 @@

+[
+    {
+        "Model": "Llama-3.1-8B-Instruct",
+        "CR": "42.6",
+        "S-Acc": "9.9",
+        "EM": "0.0",
+        "PM-0.5": "3.8",
+        "Tokens": "2478"
+    },
+    {
+        "Model": "Llama-3.1-70B-Instruct",
+        "CR": "71.8",
+        "S-Acc": "27.4",
+        "EM": "0.5",
+        "PM-0.5": "21.9",
+        "Tokens": "2090"
+    },
+    {
+        "Model": "Llama-3.3-70B-Instruct",
+        "CR": "92.4",
+        "S-Acc": "33.1",
+        "EM": "1.3",
+        "PM-0.5": "25.8",
+        "Tokens": "1842"
+    },
+    {
+        "Model": "Mistral-7B-Instruct-v0.3",
+        "CR": "85.8",
+        "S-Acc": "12.1",
+        "EM": "0.0",
+        "PM-0.5": "2.3",
+        "Tokens": "2736"
+    },
+    {
+        "Model": "Mistral-Small-Instruct-2409",
+        "CR": "91.0",
+        "S-Acc": "23.1",
+        "EM": "0.2",
+        "PM-0.5": "13.3",
+        "Tokens": "2273"
+    },
+    {
+        "Model": "Mistral-Large-Instruct-2411",
+        "CR": "96.1",
+        "S-Acc": "36.4",
+        "EM": "2.5",
+        "PM-0.5": "30.0",
+        "Tokens": "2313"
+    },
+    {
+        "Model": "Qwen2.5-7B-Instruct",
+        "CR": "85.1",
+        "S-Acc": "17.7",
+        "EM": "0.3",
+        "PM-0.5": "5.1",
+        "Tokens": "2086"
+    },
+    {
+        "Model": "Qwen2.5-32B-Instruct",
+        "CR": "96.2",
+        "S-Acc": "29.9",
+        "EM": "0.6",
+        "PM-0.5": "14.8",
+        "Tokens": "1924"
+    },
+    {
+        "Model": "Qwen2.5-72B-Instruct",
+        "CR": "95.0",
+        "S-Acc": "33.9",
+        "EM": "0.9",
+        "PM-0.5": "20.8",
+        "Tokens": "2359"
+    },
+    {
+        "Model": "QwQ-32B-Preview",
+        "CR": "65.0",
+        "S-Acc": "26.6",
+        "EM": "8.5",
+        "PM-0.5": "19.3",
+        "Tokens": "6709"
+    },
+    {
+        "Model": "DeepSeek-R1",
+        "CR": "100.0",
+        "S-Acc": "58.4",
+        "EM": "20.0",
+        "PM-0.5": "62.0",
+        "Tokens": "9856"
+    },
+    {
+        "Model": "Gemini-2.0-flash",
+        "CR": "81.1",
+        "S-Acc": "37.0",
+        "EM": "2.4",
+        "PM-0.5": "34.5",
+        "Tokens": "2637"
+    },
+    {
+        "Model": "Gemini-2.0-flash-thinking",
+        "CR": "88.2",
+        "S-Acc": "39.4",
+        "EM": "4.3",
+        "PM-0.5": "35.0",
+        "Tokens": "3725"
+    },
+    {
+        "Model": "OpenAI-gpt-4o",
+        "CR": "99.8",
+        "S-Acc": "43.7",
+        "EM": "3.2",
+        "PM-0.5": "41.7",
+        "Tokens": "1486"
+    },
+    {
+        "Model": "OpenAI-o1-mini",
+        "CR": "97.7",
+        "S-Acc": "41.3",
+        "EM": "9.1",
+        "PM-0.5": "32.8",
+        "Tokens": "9576"
+    },
+    {
+        "Model": "OpenAI-o1-preview",
+        "CR": "96.3",
+        "S-Acc": "58.7",
+        "EM": "23.6",
+        "PM-0.5": "61.7",
+        "Tokens": "11436"
+    }
+]

Sudoku.json CHANGED Viewed

@@ -13,7 +13,7 @@
         "S-Acc": "24.2",
         "EM": "1.0",
         "PM-0.5": "17.5",
-        "Tokens": "1939"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
@@ -21,7 +21,7 @@
         "S-Acc": "34.8",
         "EM": "7.0",
         "PM-0.5": "22.5",
-        "Tokens": "1061"
     },
     {
         "Model": "Mistral-7B-Instruct-v0.3",
@@ -61,7 +61,7 @@
         "S-Acc": "42.8",
         "EM": "3.5",
         "PM-0.5": "30.5",
-        "Tokens": "1201"
     },
     {
         "Model": "Qwen2.5-72B-Instruct",
@@ -85,10 +85,10 @@
         "S-Acc": "70.3",
         "EM": "50.0",
         "PM-0.5": "64.0",
-        "Tokens": "8276"
     },
     {
-        "Model": "Gemini-2.0-exp",
         "CR": "93.0",
         "S-Acc": "45.3",
         "EM": "12.5",
@@ -96,12 +96,12 @@
         "Tokens": "2842"
     },
     {
-        "Model": "Gemini-2.0-thinking",
         "CR": "79.5",
         "S-Acc": "46.5",
         "EM": "16.5",
         "PM-0.5": "41.0",
-        "Tokens": "3852"
     },
     {
         "Model": "OpenAI-gpt-4o",
@@ -109,7 +109,7 @@
         "S-Acc": "52.2",
         "EM": "14.5",
         "PM-0.5": "48.0",
-        "Tokens": "1103"
     },
     {
         "Model": "OpenAI-o1-mini",
@@ -117,7 +117,7 @@
         "S-Acc": "53.4",
         "EM": "27.0",
         "PM-0.5": "43.0",
-        "Tokens": "3960"
     },
     {
         "Model": "OpenAI-o1-preview",
@@ -125,6 +125,6 @@
         "S-Acc": "65.1",
         "EM": "50.0",
         "PM-0.5": "55.5",
-        "Tokens": "8061"
     }
 ]

         "S-Acc": "24.2",
         "EM": "1.0",
         "PM-0.5": "17.5",
+        "Tokens": "1940"
     },
     {
         "Model": "Llama-3.3-70B-Instruct",
         "S-Acc": "34.8",
         "EM": "7.0",
         "PM-0.5": "22.5",
+        "Tokens": "1062"
     },
     {
         "Model": "Mistral-7B-Instruct-v0.3",
         "S-Acc": "42.8",
         "EM": "3.5",
         "PM-0.5": "30.5",
+        "Tokens": "1202"
     },
     {
         "Model": "Qwen2.5-72B-Instruct",
         "S-Acc": "70.3",
         "EM": "50.0",
         "PM-0.5": "64.0",
+        "Tokens": "8277"
     },
     {
+        "Model": "Gemini-2.0-flash",
         "CR": "93.0",
         "S-Acc": "45.3",
         "EM": "12.5",
         "Tokens": "2842"
     },
     {
+        "Model": "Gemini-2.0-flash-thinking",
         "CR": "79.5",
         "S-Acc": "46.5",
         "EM": "16.5",
         "PM-0.5": "41.0",
+        "Tokens": "3853"
     },
     {
         "Model": "OpenAI-gpt-4o",
         "S-Acc": "52.2",
         "EM": "14.5",
         "PM-0.5": "48.0",
+        "Tokens": "1104"
     },
     {
         "Model": "OpenAI-o1-mini",
         "S-Acc": "53.4",
         "EM": "27.0",
         "PM-0.5": "43.0",
+        "Tokens": "3961"
     },
     {
         "Model": "OpenAI-o1-preview",
         "S-Acc": "65.1",
         "EM": "50.0",
         "PM-0.5": "55.5",
+        "Tokens": "8062"
     }
 ]

index.html CHANGED Viewed

@@ -203,6 +203,7 @@
         document.getElementById("defaultOpen").click();
         const urls = {
             'Acrostic': 'Acrostic.json',
             'Drop Quote': 'Drop_Quote.json',
             'Crossword': 'Crossword.json',

         document.getElementById("defaultOpen").click();
         const urls = {
+            'Overall': 'Overall.json',
             'Acrostic': 'Acrostic.json',
             'Drop Quote': 'Drop_Quote.json',
             'Crossword': 'Crossword.json',