1import pyarrow as pa
2
3schema = pa.schema([
4    # 图片列：large_binary + blob 编码 + 模态声明
5    pa.field("image", pa.large_binary(), metadata={
6        "bos-dataset-field:modality": "image",
7        "bos-dataset-field:mimetype": "image/jpeg",  # RFC 类型，照实填写
8        "bos-dataset-field:format": "jpg",           # 封装格式，照实填写
9        "lance-encoding:blob": "true"}),             # 大对象用 blob 编码
10    # 音频列
11    pa.field("audio", pa.large_binary(), metadata={
12        "bos-dataset-field:modality": "audio",
13        "bos-dataset-field:mimetype": "audio/wav",
14        "bos-dataset-field:format": "wav",
15        "lance-encoding:blob": "true"}),
16    # 视频列
17    pa.field("video", pa.large_binary(), metadata={
18        "bos-dataset-field:modality": "video",
19        "bos-dataset-field:mimetype": "video/mp4",
20        "bos-dataset-field:format": "mp4",
21        "lance-encoding:blob": "true"}),
22    # 文本列
23    pa.field("caption", pa.string(),
24             metadata={"bos-dataset-field:modality": "text"}),
25    # 普通元数据列可不声明；未知键会被忽略，不影响读写
26    pa.field("label", pa.string()),
27])
28
29table = pa.Table.from_pylist([...], schema=schema)  # 数据按 schema 顺序放进
30lance.write_dataset(table, uri, storage_options=so)

1SELECT * FROM dataset LIMIT 10;

1-- 按标签统计行数
2SELECT label, COUNT(*) AS cnt FROM dataset GROUP BY label;
3-- 查看训练/测试切分比例
4SELECT split, COUNT(*) AS cnt FROM dataset GROUP BY split;