← back
arXivYunpeng Xu, Kun ZhengTue, Sep 8, 2026, 10:31 AM PDT
score 17.1

AI training data mix: each subject area has a best share, and later fine-tuning can't fix a poor mix

Original: Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training

Source: arxiv.org

Writing ELI5 summary…

AI training data mix: each subject area has a best share, and later fine-tuning can't fix a poor mix · TinyNews · TinyNews