SDIR: An Open Tool to Catch Training Data Collapsing Before You Train

2026-08-03Medium英文实操工具

Summary

HAS-Core 框架首个实操协议 SDIR（合成数据近交率）：训练前读取训练批次结构，检测数据源收缩、递归合成与语义多样性损失；只读不修改用户代码。

Read the original ↗

Original URLs: https://medium.com/@huqingyun/sdir-an-open-tool-to-catch-training-data-collapsing-before-you-train-b85e71b4427f · https://doi.org/10.5281/zenodo.21778972 · https://doi.org/10.5281/zenodo.21813259

Citation: Hu-Yang, Qingyun. “SDIR: An Open Tool to Catch Training Data Collapsing Before You Train.” 2026-08-03. harmondeg.org archive page a93.