Spaces:
Running
Running
Your Name commited on
Commit ยท
276df52
1
Parent(s): ad59b21
update gpt 5
Browse files- .DS_Store +0 -0
- assets/model_color.json +2 -0
- rank_data_03_25_2025.json +33 -0
.DS_Store
CHANGED
|
Binary files a/.DS_Store and b/.DS_Store differ
|
|
|
assets/model_color.json
CHANGED
|
@@ -16,6 +16,7 @@
|
|
| 16 |
"gpt-4o-2024-11-20": "#00BFA5",
|
| 17 |
"gpt-4.5-preview-2025-02-27": "#00796B",
|
| 18 |
"gpt-4.1-2025-04-14": "#00897B",
|
|
|
|
| 19 |
"gpt-oss-20b": "#4DD0E1",
|
| 20 |
"gpt-oss-120b": "#00838F",
|
| 21 |
|
|
@@ -58,6 +59,7 @@
|
|
| 58 |
"๐ฎ gpt-4o-2024-11-20": "#00BFA5",
|
| 59 |
"๐ฎ gpt-4.5-preview-2025-02-27": "#00796B",
|
| 60 |
"๐ฎ gpt-4.1-2025-04-14": "#00897B",
|
|
|
|
| 61 |
"๐ฎ gpt-oss-20b": "#4DD0E1",
|
| 62 |
"๐ฎ gpt-oss-120b": "#00838F",
|
| 63 |
|
|
|
|
| 16 |
"gpt-4o-2024-11-20": "#00BFA5",
|
| 17 |
"gpt-4.5-preview-2025-02-27": "#00796B",
|
| 18 |
"gpt-4.1-2025-04-14": "#00897B",
|
| 19 |
+
"gpt-5-thinking-high": "#00695C",
|
| 20 |
"gpt-oss-20b": "#4DD0E1",
|
| 21 |
"gpt-oss-120b": "#00838F",
|
| 22 |
|
|
|
|
| 59 |
"๐ฎ gpt-4o-2024-11-20": "#00BFA5",
|
| 60 |
"๐ฎ gpt-4.5-preview-2025-02-27": "#00796B",
|
| 61 |
"๐ฎ gpt-4.1-2025-04-14": "#00897B",
|
| 62 |
+
"๐ฎ gpt-5-thinking-high": "#00695C",
|
| 63 |
"๐ฎ gpt-oss-20b": "#4DD0E1",
|
| 64 |
"๐ฎ gpt-oss-120b": "#00838F",
|
| 65 |
|
rank_data_03_25_2025.json
CHANGED
|
@@ -67,6 +67,12 @@
|
|
| 67 |
"score": 986.97,
|
| 68 |
"detail_data":"986.97",
|
| 69 |
"progress": "1-1"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 70 |
}
|
| 71 |
]
|
| 72 |
},
|
|
@@ -216,6 +222,12 @@
|
|
| 216 |
"score": 2040.0,
|
| 217 |
"details": "1484, 1432, 3204",
|
| 218 |
"highest_tail": 256
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 219 |
}
|
| 220 |
]
|
| 221 |
},
|
|
@@ -341,6 +353,11 @@
|
|
| 341 |
"model": "๐ฎ gpt-oss-20b",
|
| 342 |
"score": 14.33,
|
| 343 |
"details": "14, 13, 16"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 344 |
}
|
| 345 |
]
|
| 346 |
},
|
|
@@ -466,6 +483,11 @@
|
|
| 466 |
"model": "๐ฎ gpt-oss-20b",
|
| 467 |
"score": 249.67,
|
| 468 |
"details": "243, 262, 244"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 469 |
}
|
| 470 |
]
|
| 471 |
},
|
|
@@ -615,6 +637,12 @@
|
|
| 615 |
"score": 4,
|
| 616 |
"detail_box_on_target": "4",
|
| 617 |
"cracked_levels": "2"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 618 |
}
|
| 619 |
]
|
| 620 |
},
|
|
@@ -719,6 +747,11 @@
|
|
| 719 |
"model": "๐ฎ gemini-2.5-flash-preview-05-20",
|
| 720 |
"score": 4.33,
|
| 721 |
"details": "3,4,6"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 722 |
}
|
| 723 |
]
|
| 724 |
}
|
|
|
|
| 67 |
"score": 986.97,
|
| 68 |
"detail_data":"986.97",
|
| 69 |
"progress": "1-1"
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"model": "๐ฎ gpt-5-thinking-high",
|
| 73 |
+
"score": 1432.33,
|
| 74 |
+
"detail_data":"1875.0, 1397.0, 1025.0",
|
| 75 |
+
"progress": "1-1"
|
| 76 |
}
|
| 77 |
]
|
| 78 |
},
|
|
|
|
| 222 |
"score": 2040.0,
|
| 223 |
"details": "1484, 1432, 3204",
|
| 224 |
"highest_tail": 256
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"model": "๐ฎ gpt-5-thinking-high",
|
| 228 |
+
"score": 5770.67,
|
| 229 |
+
"detail_data":"4516, 8116, 4680",
|
| 230 |
+
"highest_tail": 512
|
| 231 |
}
|
| 232 |
]
|
| 233 |
},
|
|
|
|
| 353 |
"model": "๐ฎ gpt-oss-20b",
|
| 354 |
"score": 14.33,
|
| 355 |
"details": "14, 13, 16"
|
| 356 |
+
},
|
| 357 |
+
{
|
| 358 |
+
"model": "๐ฎ gpt-5-thinking-high",
|
| 359 |
+
"score": 84.33,
|
| 360 |
+
"details": "36, 112, 105"
|
| 361 |
}
|
| 362 |
]
|
| 363 |
},
|
|
|
|
| 483 |
"model": "๐ฎ gpt-oss-20b",
|
| 484 |
"score": 249.67,
|
| 485 |
"details": "243, 262, 244"
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"model": "๐ฎ gpt-5-thinking-high",
|
| 489 |
+
"score": 615.67,
|
| 490 |
+
"details": "573, 694, 580"
|
| 491 |
}
|
| 492 |
]
|
| 493 |
},
|
|
|
|
| 637 |
"score": 4,
|
| 638 |
"detail_box_on_target": "4",
|
| 639 |
"cracked_levels": "2"
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"model": "๐ฎ gpt-5-thinking-high",
|
| 643 |
+
"score": 11,
|
| 644 |
+
"detail_box_on_target": "11, 11",
|
| 645 |
+
"cracked_levels": "5"
|
| 646 |
}
|
| 647 |
]
|
| 648 |
},
|
|
|
|
| 747 |
"model": "๐ฎ gemini-2.5-flash-preview-05-20",
|
| 748 |
"score": 4.33,
|
| 749 |
"details": "3,4,6"
|
| 750 |
+
},
|
| 751 |
+
{
|
| 752 |
+
"model": "๐ฎ gpt-5-thinking-high",
|
| 753 |
+
"score": 9,
|
| 754 |
+
"details": "9,9"
|
| 755 |
}
|
| 756 |
]
|
| 757 |
}
|