Querit-Reranker: Training Compact Multilingual Rerankers via Efficient Label-Free Distribution Adaptation
Researchers have developed Querit-Reranker, a new family of multilingual cross-encoder rerankers designed for efficient adaptation to various ranking tasks without requiring extensive labeled data. The models are trained using a pipeline that leverages synthetic query mining and teacher scores as soft labels, and checkpoints can be merged to create a single deployable model. Querit-Reranker-A0.4B demonstrated significant improvements on benchmarks like BEIR and MIRACL, while Querit-Reranker-4B achieved state-of-the-art performance among publicly available models. Both models are available on Hugging Face. AI
IMPACT Introduces a more efficient method for adapting multilingual rerankers, potentially lowering the barrier for deploying advanced search and retrieval systems.