NHNM mines natural hard negatives for instruction retrieval
Original title: Natural Hard-Negative Mining from Positives for Instruction-Following Retrieval
Fine-tuning 568M to 4B models on nearly 7K NHNM-MARCO instances matches or beats much larger retrievers on InstructIR and FollowIR.
Natural Hard-Negative Mining builds instruction-following retrieval supervision by turning original positives into hard negatives under query-compatible instructions. An automated pipeline of Candidate Cleaning, Reversal Construction, and Geometry Filtering produces compact contrastive data aligned with the pretrained retrieval space. Applied to MSMARCO, it yields NHNM-MARCO, nearly 7K training instances. Models from 568M to 4B parameters fine-tuned on that set show strong results on InstructIR and FollowIR and match or surpass substantially larger retrievers in several settings. Ablations attribute the gains to the hard-negative signal rather than scale or data volume.
Unconfirmed: Performance claims are the authors' reported comparisons; no independent evaluation is stated.
- NHNM-MARCO contains nearly 7K instruction-following instances.
- Evaluated models range from 568M to 4B parameters.
- Benchmarks are InstructIR and FollowIR.
- Ablations credit natural hard negatives over scale or data volume.
From the original article
1 min read at arxiv.orgInstruction-following information retrieval (IFIR) requires retrievers to rank documents by both a query and its instance-specific instructions, yet instruction-conditioned contrastive supervision remains scarce. Existing synthetic methods typically fix a positive passage and generate additional negatives, overlooking the hard-negative potential of the original positive.
By Chaolang Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.