late-interaction
ro-retrieval-triplets-late-interaction
Romanian retrieval triplets (late-interaction hard negatives)
(anchor, positive, negative) triplets for training a Romanian retriever,
generated with Gemma 3 27B over a merged Romanian corpus and mined for hard
negatives with late interaction (MaxSim).
A BM25 variant of the same triplets is at
PaulBurca2005/ro-retrieval-triplets.
The two are close to disjoint - over the queries both contain, mean Jaccard
overlap between their negative sets is 0.053 and 73% share no negative at… See the full description on the dataset page: https://huggingface.co/datasets/PaulBurca2005/ro-retrieval-triplets-late-interaction.hybrid-late-interaction-bm25
When Does Lexical Search Help Late-Interaction Retrieval?
This repository contains the paper, code, configurations, and result artifacts for a study of lexical+semantic candidate generation for late-interaction reranking.
The central finding is that BM25 helps late-interaction retrieval most as part of a hybrid first-stage candidate generator. On seven BEIR-style datasets, ModernColBERT reranking over lexical+semantic RRF top-100 candidates improves average nDCG@10 from 0.3826… See the full description on the dataset page: https://huggingface.co/datasets/datamokotow/hybrid-late-interaction-bm25.