Understanding SFT Mechanisms in LLMs: Resolving Practice Disputes and Avoiding Wasted Compute

吃透大模型SFT底层机理：终结实践争议，规避无效算力

Junpeng Zhang and coauthors argue that SFT on highly homogeneous data has an effective window of only hundreds to about 1,000 training steps, and their interaction-based warning signal detects overfitting before loss gaps appear, saving roughly 30%–50% of training compute.

Why it matters: HKR-H/K/R all pass: the paper gives testable SFT windows, earlier overfitting warnings, and 30%-50% compute savings. It is strong research, not a major model or product release, so it stays below 85.