yoshinakachi commited on
Commit
493ead8
·
verified ·
1 Parent(s): 5b4f4bd

Update src/App.js

Browse files
Files changed (1) hide show
  1. src/App.js +43 -43
src/App.js CHANGED
@@ -1,93 +1,93 @@
1
  import React, { useState, useEffect } from 'react';
2
 
3
  const BenchmarkChart = () => {
4
- // Real data sorted by Human Baseline (ascending)
5
  const benchmarkData = [
6
  {
7
- model: "GPT 5 mini",
8
- baseline: 7.67,
9
  methods: {
10
- keyword_objective_combined: 11.28,
11
- root_problem: 10.44
12
  }
13
  },
14
  {
15
- model: "GPT 5",
16
- baseline: 8.33,
17
  methods: {
18
- keyword_objective_combined: 11.68,
19
- root_problem: 12.46
20
  }
21
  },
22
  {
23
- model: "Claude Opus 4.1",
24
- baseline: 20.67,
25
  methods: {
26
- keyword_objective_combined: 23.56
27
  }
28
  },
29
  {
30
- model: "GPT o3",
31
- baseline: 22.00,
32
  methods: {
33
- keyword_objective_combined: 30.53
 
34
  }
35
  },
36
  {
37
- model: "Claude 4 Sonnet",
38
- baseline: 26.33,
39
  methods: {
40
- keyword_objective_combined: 28.64
41
  }
42
  },
43
  {
44
- model: "Llama 4 Maverick Instruct",
45
- baseline: 45.00,
46
  methods: {
47
- keyword_objective_combined: 56.46
48
  }
49
  },
50
  {
51
- model: "Mixtral 8x22B",
52
- baseline: 48.00,
53
  methods: {
54
- keyword_objective_combined: 66.82
55
  }
56
  },
57
  {
58
- model: "Gemini 2.5 Pro",
59
- baseline: 55.67,
60
  methods: {
61
- keyword_objective_combined: 74.14,
62
- root_problem: 67.19
63
  }
64
  },
65
  {
66
- model: "Llama 3 8B Instruct Reference",
67
- baseline: 58.33,
68
  methods: {
69
- keyword_objective_combined: 68.86
70
  }
71
  },
72
  {
73
- model: "Llama 3.1 405B",
74
- baseline: 67.00,
75
  methods: {
76
- keyword_objective_combined: 80.75
77
  }
78
  },
79
  {
80
- model: "Deepseek R1-0528",
81
- baseline: 68.67,
82
  methods: {
83
- keyword_objective_combined: 83.76
 
84
  }
85
  },
86
  {
87
- model: "Grok 4",
88
- baseline: 68.67,
89
  methods: {
90
- keyword_objective_combined: 85.15
 
91
  }
92
  }
93
  ];
@@ -175,9 +175,9 @@ const BenchmarkChart = () => {
175
  </div>
176
  </div>
177
 
178
- {/* Chart Container - Taller */}
179
  <div className="bg-white rounded-xl shadow-2xl p-4">
180
- <div className="h-[600px] overflow-y-auto pr-2">
181
  <div className="space-y-2">
182
  {benchmarkData.map((modelData, index) => {
183
  const currentValue = getCurrentValue(modelData, currentPhase);
@@ -274,7 +274,7 @@ const BenchmarkChart = () => {
274
  {/* Footer */}
275
  <div className="mt-4 text-center text-slate-400 space-y-1">
276
  <p className="text-sm">
277
- Sorted by Human Baseline ASR • Click button above for manual control
278
  </p>
279
  <p className="text-xs">
280
  Bars extend from baseline to show transformation method impact
 
1
  import React, { useState, useEffect } from 'react';
2
 
3
  const BenchmarkChart = () => {
4
+ // Real data sorted by highest achievable ASR (largest bars at top)
5
  const benchmarkData = [
6
  {
7
+ model: "Grok 4",
8
+ baseline: 68.67,
9
  methods: {
10
+ keyword_objective_combined: 85.15
 
11
  }
12
  },
13
  {
14
+ model: "Deepseek R1-0528",
15
+ baseline: 68.67,
16
  methods: {
17
+ keyword_objective_combined: 83.76
 
18
  }
19
  },
20
  {
21
+ model: "Llama 3.1 405B",
22
+ baseline: 67.00,
23
  methods: {
24
+ keyword_objective_combined: 80.75
25
  }
26
  },
27
  {
28
+ model: "Gemini 2.5 Pro",
29
+ baseline: 55.67,
30
  methods: {
31
+ keyword_objective_combined: 74.14,
32
+ root_problem: 67.19
33
  }
34
  },
35
  {
36
+ model: "Llama 3 8B Instruct Reference",
37
+ baseline: 58.33,
38
  methods: {
39
+ keyword_objective_combined: 68.86
40
  }
41
  },
42
  {
43
+ model: "Mixtral 8x22B",
44
+ baseline: 48.00,
45
  methods: {
46
+ keyword_objective_combined: 66.82
47
  }
48
  },
49
  {
50
+ model: "Llama 4 Maverick Instruct",
51
+ baseline: 45.00,
52
  methods: {
53
+ keyword_objective_combined: 56.46
54
  }
55
  },
56
  {
57
+ model: "GPT o3",
58
+ baseline: 22.00,
59
  methods: {
60
+ keyword_objective_combined: 30.53
 
61
  }
62
  },
63
  {
64
+ model: "Claude 4 Sonnet",
65
+ baseline: 26.33,
66
  methods: {
67
+ keyword_objective_combined: 28.64
68
  }
69
  },
70
  {
71
+ model: "Claude Opus 4.1",
72
+ baseline: 20.67,
73
  methods: {
74
+ keyword_objective_combined: 23.56
75
  }
76
  },
77
  {
78
+ model: "GPT 5",
79
+ baseline: 8.33,
80
  methods: {
81
+ keyword_objective_combined: 11.68,
82
+ root_problem: 12.46
83
  }
84
  },
85
  {
86
+ model: "GPT 5 mini",
87
+ baseline: 7.67,
88
  methods: {
89
+ keyword_objective_combined: 11.28,
90
+ root_problem: 10.44
91
  }
92
  }
93
  ];
 
175
  </div>
176
  </div>
177
 
178
+ {/* Chart Container - Longer */}
179
  <div className="bg-white rounded-xl shadow-2xl p-4">
180
+ <div className="h-[700px] overflow-y-auto pr-2">
181
  <div className="space-y-2">
182
  {benchmarkData.map((modelData, index) => {
183
  const currentValue = getCurrentValue(modelData, currentPhase);
 
274
  {/* Footer */}
275
  <div className="mt-4 text-center text-slate-400 space-y-1">
276
  <p className="text-sm">
277
+ Sorted by Maximum ASR (largest bars at top) • Click button above for manual control
278
  </p>
279
  <p className="text-xs">
280
  Bars extend from baseline to show transformation method impact