| --- |
| pretty_name: Multi-kind Synthetic Dataset Lesson |
| license: apache-2.0 |
| language: |
| - zh |
| task_categories: |
| - time-series-forecasting |
| - tabular-classification |
| - text-generation |
| tags: |
| - synthetic |
| - tutorial |
| - multi-modal |
| size_categories: |
| - 10K<n<100K |
| configs: |
| - config_name: timeseries |
| default: true |
| data_files: |
| - split: train |
| path: data/timeseries/train.arrow |
| - split: validation |
| path: data/timeseries/validation.arrow |
| - split: test |
| path: data/timeseries/test.arrow |
| - config_name: tabular |
| data_files: |
| - split: train |
| path: data/tabular/train.arrow |
| - split: validation |
| path: data/tabular/validation.arrow |
| - split: test |
| path: data/tabular/test.arrow |
| - config_name: text |
| data_files: |
| - split: train |
| path: data/text/train.arrow |
| - split: validation |
| path: data/text/validation.arrow |
| - split: test |
| path: data/text/test.arrow |
| --- |
| |
| # 多类型合成数据练习 |
|
|
| 全部由脚本生成，包含时序、表格、文本三类数据。每类都有 train、validation、test。 |
| configs 统一读取 data/ 下的 Arrow stream 文件；raw/ 保留 CSV 和 JSONL 原始文件。 |
|
|
| | 配置 | train | validation | test | 字段 | |
| |---|---:|---:|---:|---| |
| | timeseries（默认） | 3 | 3 | 3 | item_id, start, end, freq, target | |
| | tabular | 12 | 4 | 4 | sample_id, temperature, humidity, label | |
| | text | 12 | 4 | 4 | sample_id, text, label | |
| |
| 时序每行是一台传感器的一段数据，三段分别为前 144、后 48、最后 48 小时。 |
| 同一传感器出现在各 split，但时间点不重叠；这是预测已知传感器未来值的评估。 |
| 表格和文本是独立合成样本，按种子 42 打乱后按 60%/20%/20% 划分，sample_id 不重叠。 |
| 样例不适用于真实业务结论；重复模板文本只用于学习文件与划分流程。 |
|
|
| ```python |
| from datasets import load_dataset |
| data = load_dataset("YOUR_USERNAME/multi-data-demo", name="text") |
| print(data["train"][0]) |
| ``` |
|
|
