File size: 5,446 Bytes
08818c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
<!DOCTYPE html>
<html lang="en">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <style>
        body {
            background-color: #111;
            font-family: Arial, sans-serif;
            color: #fff;
            display: flex;
            justify-content: center;
            align-items: center;
            flex-direction: column;
            height: 100vh;
            margin: 0;
        }

        h1 {
            font-size: 36px;
            margin-bottom: 10px;
        }

        h2 {
            font-size: 18px;
            font-weight: normal;
            margin-bottom: 30px;
            color: #ccc;
        }

        table {
            width: 90%;
            border-collapse: separate;
            border-spacing: 0;
            background-color: #1b1b1b;
            border-radius: 12px;
            overflow: hidden;
            margin: 20px 0;
            table-layout: fixed;
        }

        th, td {
            text-align: center;
            padding: 12px;
            border: 1px solid #333;
            vertical-align: middle;
        }

        th {
            background-color: #222;
            font-weight: bold;
            font-size: 14px;
        }

        td {
            background-color: #1b1b1b;
            font-size: 14px;
            word-wrap: break-word;
        }

        .highlight-column {
            border-left: 3px solid #0066ff;
            border-right: 3px solid #0066ff;
        }

        .highlight-header {
            border-top: 3px solid #0066ff;
            border-top-left-radius: 12px;
            border-top-right-radius: 12px;
        }

        .highlight-footer {
            border-bottom: 3px solid #0066ff;
            border-bottom-left-radius: 12px;
            border-bottom-right-radius: 12px;
        }

        .bold {
            font-weight: 900; /* Extra bold */
        }

        tr:first-child th:first-child {
            border-top-left-radius: 12px;
        }

        tr:first-child th:last-child {
            border-top-right-radius: 12px;
        }

        tr:last-child td:first-child {
            border-bottom-left-radius: 12px;
        }

        tr:last-child td:last-child {
            border-bottom-right-radius: 12px;
        }

        .footnote {
            font-size: 12px;
            color: #888;
            text-align: left;
            max-width: 90%;
            margin-top: 20px;
        }

    </style>
</head>
<body>

<h1>田忌赛马</h1>
<h2>Goodhart's Law on Benchmarks</h2>

<table>
    <tr>
        <th>Capability</th>
        <th>Description</th>
        <th class="highlight-column highlight-header">miniG</th>
        <th>Gemini-Flash</th>
        <th>GLM-4-9B-Chat</th>
        <th>Llama 3.1 8B Instruct</th>
    </tr>
    <tr>
        <td class="bold">MMLU</td>
        <td>Representation of questions in 57 subjects<br>(incl. STEM, humanities, and others)</td>
        <td class="highlight-column bold">85.45</td>
        <td>78.9</td>
        <td>72.4</td>
        <td>69.4</td>
    </tr>
    <tr>
        <td class="bold">IFEval</td>
        <td>Evaluation of instruction-following<br>using verifiable prompts</td>
        <td class="highlight-column">74.22</td>
        <td>-</td>
        <td>69</td>
        <td class="bold">80.4</td>
    </tr>
    <tr>
        <td class="bold">GSM8K</td>
        <td>Challenging math problems<br>(5-shot evaluation)</td>
        <td class="highlight-column">75.89 (5-shot)</td>
        <td class="bold">86.2 (11-shot)</td>
        <td>79.6</td>
        <td>84.5 (8-shot CoT)</td>
    </tr>
    <tr>
        <td class="bold">HumanEval</td>
        <td>Python code generation on a held-out dataset<br>(0-shot)</td>
        <td class="highlight-column bold">79.88</td>
        <td>74.3</td>
        <td>71.8</td>
        <td>72.6</td>
    </tr>
    <tr>
        <td class="bold">GPQA</td>
        <td>Challenging dataset of questions<br>from biology, physics, and chemistry</td>
        <td class="highlight-column">37.37</td>
        <td class="bold">39.5</td>
        <td>34.3 (base)</td>
        <td>34.2</td>
    </tr>
    <tr>
        <td class="bold">Context Window</td>
        <td>Maximum context length<br>the model can handle</td>
        <td class="highlight-column bold">1M</td>
        <td class="bold">1M</td>
        <td>128K</td>
        <td>128K</td>
    </tr>
    <tr>
        <td class="bold">Input</td>
        <td>Supported input modalities</td>
        <td class="highlight-column highlight-footer">Text, image<br>(single model)</td>
        <td>Text, image, audio, video</td>
        <td>Text only</td>
        <td>Text only</td>
    </tr>
</table>

<div class="footnote">
    1. miniG is a 14B parameter model derived from the 9B parameter glm-4-9b-chat-1m model weights. It continues pre-training on a selected corpus of 20B tokens while retaining long-context capabilities. The model is fine-tuned on a dataset of 120M+ conversation entries, synthesized through cross-page clustering similar to RAG on this selected corpus. Additionally, miniG underwent multimodal training in two stages for single image input, with the second stage reinitializing 5B parameters of a Vision Transformer from glm-4v-9b for Locked-Image Tuning.<br>
    2. miniG outputs are formatted similarly to Gemini 1.5 Flash but were not trained on data generated by the Gemini models.
</div>

</body>
</html>