Tom Aarsen released six new CrossEncoder reranker models built on Ettin ModernBERT encoders, achieving state-of-the-art performance at their respective sizes. The models were trained using a distillation recipe based on Mixedbread AI's mxbai-rerank-large-v2 scores over the cross-encoder/ettin-reranker-v1-data dataset, which combines subsets from Lighton's embedding datasets. Evaluated on MTEB(eng, v2) Retrieval paired with Google's google/embeddinggemma-300m, the rerankers are available on Hugging Face with full training data and recipes.
No score is assigned. Sources and their independence are shown in the citation chain below.