added final notes about mining

README.md

@@ -27,6 +27,8 @@ The negative samples have been automatically mined with static bilingual [sts-mr
     sampling_strategy: "top", # "top" means that we sample the top candidates as negatives
     num_negatives: 5, i.e. we mine 5 negatives per question-answer pair
 
+The 'relative_margin' below the hit of 5% as 'safe space' is inspired by the [NV-Retriever paper](https://arxiv.org/abs/2407.15831) - the same as explained in the documentation of [mine_hard_negatives](https://sbert.net/docs/package_reference/util.html#sentence_transformers.util.mine_hard_negatives). But I wrote my own miner based upon [USearch](https://github.com/unum-cloud/usearch) only.
+
 ### Dropped lines without hard negatives
 Almost forgot to rescue: The remaining 623_000 deduplicated lines of material so unique that no negatives could get mined have been rescued into a seperate dataset: [gooaq_mt_german_0_hard_negatives](https://huggingface.co/datasets/MarcGrumpyOlejak/gooaq_mt_german_0_hard_negatives)