Natural Hard-Negative Mining from Positives for Instruction-Following Retrieval

Natural Hard-Negative Mining (NHNM) repurposes positive passages as instruction-specific hard negatives to address data scarcity in instruction-following retrieval. The pipeline uses candidate cleaning and geometry filtering to generate high-contrast supervision. Fine-tuning models from 568M to 4B parameters on the resulting 7K NHNM-MARCO instances improves performance on InstructIR and FollowIR benchmarks.