Natural Hard-Negative Mining from Positives for Instruction-Following Retrieval
The proposed Natural Hard-Negative Mining (NHNM), which promotes convertible negatives under query-compatible instructions and repurposes original positives as natural instruction hard negatives, produces compact, high-contrast supervision aligned with the pretrained retrieval space.
Chao-Lang Li, Jing-Yu Li, Zhao-Cheng Du et al.
· 0 citations