msft-mixture-overfitting

【数据与研究】Identify three ranked findings on multi-task SFT: (1) heterogeneous overfitting—sub-datasets peak at different training points (contradicts uniform duration practice); (2) parameter divergence—excluding 1/10 of data shifts optimal points 0.91 epochs for remaining tasks; (3) SFT compute negligible (0.01% of training). Implement mSFT: iterative roll-ou…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准