Datasets:

---

prdev

/

jxm-nomic-hard-negatives

like 1

Hi Jonah,

This dataset was generated by following the mining methodology outlined in Nv-Retriever paper (https://arxiv.org/pdf/2407.15831). I started from a subset of jxm/nomic_embed_unsupervised. To prevent contamination on the MTEB, some datasets of the nomic unsupervised set are removed to create a clean subset. From there, I applied Nv-Retriever's TopK-MarginPos method with a 5% margin.

Hope this helped!
-- Pranav

Awesome, thanks for the detailed explanation.

Appreciate it!
Jonah