cm4-synthetic-testing

This dataset is designed to be used in testing multimodal text/image models. It's derived from cm4-10k dataset.

The current splits are: `['100.unique', '100.repeat', '300.unique', '300.repeat', '1k.unique', '1k.repeat', '10k.unique', '10k.repeat']`.

The `unique` ones ensure uniqueness across text entries.

The `repeat` ones are repeating the same 10 unique records: - these are useful for memory leaks debugging as the records are always the same and thus remove the record variation from the equation.

The default split is `100.unique`.

The full process of this dataset creation is documented inside [cm4-synthetic-testing.py](./cm4-synthetic-testing.py).

本数据集专为测试多模态文本-图像模型（multimodal text/image models）打造，其衍生自cm4-10k数据集。

当前预设的数据集划分方案包括：`['100.unique', '100.repeat', '300.unique', '300.repeat', '1k.unique', '1k.repeat', '10k.unique', '10k.repeat']`。

其中带有`unique`标识的划分集可确保所有文本条目均具备唯一性。

带有`repeat`标识的划分集则重复使用10条唯一记录：此类划分集适用于内存泄漏调试，由于记录始终保持一致，能够排除记录变量对实验结果的干扰。

默认数据集划分方案为`100.unique`。

本数据集的完整构建流程已在[cm4-synthetic-testing.py](./cm4-synthetic-testing.py)文件中进行了详细说明。