File size: 5,968 Bytes
fdb108a
b0f810c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fdb108a
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
<!doctype html>
<html lang="en">
  <head>
    <meta charset="utf-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1" />
    <title>Kumo Tabular in the Browser</title>
    <meta name="description" content="Zero-shot tabular predictions on your own CSV or Parquet file with NVIDIA Kumo Tabular, DuckDB-WASM and WebGPU. Nothing leaves your browser." />
    <link rel="icon" href="data:image/svg+xml,%3Csvg xmlns=%22http://www.w3.org/2000/svg%22 viewBox=%220 0 16 16%22%3E%3Crect width=%2216%22 height=%2216%22 rx=%223%22 fill=%22%2376b900%22/%3E%3Cpath d=%22M4 4h8M4 8h8M4 12h8%22 stroke=%22white%22 stroke-width=%221.5%22/%3E%3C/svg%3E" />
    <script type="module" crossorigin src="/assets/index-1X7aveGr.js"></script>
    <link rel="stylesheet" crossorigin href="/assets/index-BXYgCNdU.css">
  </head>
  <body>
    <header class="top">
      <div class="brand">
        <h1>Kumo Tabular <span>in the browser</span></h1>
        <p>
          Drop a CSV or Parquet file, pick a column, get predictions. NVIDIA's
          <a href="https://huggingface.co/nvidia/Kumo-Tabular" target="_blank" rel="noreferrer">Kumo Tabular</a>
          learns from your labeled rows in one forward pass: no training, no server. Your data stays on this page.
        </p>
      </div>
      <div class="badges">
        <span id="device-badge" class="badge">checking GPU…</span>
        <span class="badge">DuckDB-WASM</span>
        <span class="badge">Small · 28M</span>
      </div>
    </header>

    <main>
      <section class="card" id="data-card">
        <div class="card-head">
          <h2><span class="step">1</span> Data</h2>
          <span id="data-summary" class="muted"></span>
        </div>
        <label id="drop" class="drop">
          <input id="file" type="file" accept=".csv,.parquet,.tsv,.txt" hidden />
          <strong>Drop a CSV or Parquet file</strong>
          <span class="muted">or click to choose one. It is read locally by DuckDB.</span>
        </label>
        <div class="samples">
          <span class="muted">Or try:</span>
          <button data-sample="churn_synthetic.csv" data-target="churn">Churn (150 unlabeled rows)</button>
          <button data-sample="breast_cancer.csv" data-target="diagnosis">Breast cancer</button>
          <button data-sample="iris.csv" data-target="species">Iris</button>
          <button data-sample="california_housing.csv" data-target="MedHouseVal">California housing</button>
          <button data-sample="diabetes.csv" data-target="progression">Diabetes</button>
        </div>
        <details id="sql-details">
          <summary>Filter or reshape with SQL</summary>
          <p class="muted">The model sees the result of this query. The file is table <code>source</code>.</p>
          <textarea id="sql" spellcheck="false">SELECT * FROM source</textarea>
          <button id="apply-sql" class="secondary">Apply</button>
        </details>
      </section>

      <section class="card" id="setup-card" hidden>
        <div class="card-head">
          <h2><span class="step">2</span> What to predict</h2>
        </div>
        <div class="setup">
          <label>Target column <select id="target"></select></label>
          <label>Task
            <select id="task">
              <option value="classification">Classification (up to 10 classes)</option>
              <option value="regression">Regression</option>
            </select>
          </label>
          <label>Rows to predict
            <select id="mode">
              <option value="missing">Rows where the target is empty</option>
              <option value="holdout">Hold out 20% and score them</option>
            </select>
          </label>
        </div>
        <p id="split-note" class="muted"></p>
        <div class="columns-head">
          <span>Features</span>
          <span class="muted">numbers are numerical; text and booleans are categorical</span>
        </div>
        <div id="columns" class="columns"></div>
        <div class="run-row">
          <button id="run" class="primary">Predict</button>
          <div id="progress" class="progress" hidden><div></div></div>
          <span id="status" class="muted"></span>
        </div>
      </section>

      <section class="card" id="results-card" hidden>
        <div class="card-head">
          <h2><span class="step">3</span> Predictions</h2>
          <span id="timing" class="muted"></span>
        </div>
        <div id="metrics" class="metrics"></div>
        <div class="results-grid">
          <div id="chart" class="chart"></div>
          <div class="table-wrap"><table id="preview"></table></div>
        </div>
        <details open>
          <summary>Query the predictions with SQL</summary>
          <p class="muted">Tables: <code>source</code>, <code>data</code>, <code>predictions</code>, and <code>results</code> (data joined with predictions).</p>
          <textarea id="query" spellcheck="false"></textarea>
          <div class="run-row">
            <button id="run-query" class="secondary">Run query</button>
            <button id="download" class="secondary">Download results.csv</button>
          </div>
          <div class="table-wrap"><table id="query-out"></table></div>
        </details>
      </section>
    </main>

    <footer>
      Model: <a href="https://huggingface.co/nvidia/Kumo-Tabular" target="_blank" rel="noreferrer">nvidia/Kumo-Tabular</a>
      (Small, OpenMDW-1.1), converted to ONNX. Preprocessing ports
      <a href="https://github.com/NVIDIA/structured-data-models" target="_blank" rel="noreferrer">structured-data-models</a>'
      default recipe (one estimator) and matches it value for value. Runs with
      <a href="https://huggingface.co/docs/transformers.js" target="_blank" rel="noreferrer">Transformers.js</a> and
      <a href="https://duckdb.org/docs/api/wasm/overview" target="_blank" rel="noreferrer">DuckDB-WASM</a>.
    </footer>
  </body>
</html>