Japanese Reranker XSmall V2

hotchpotch/japanese-reranker-xsmall-v2 is a reranker model that scores query–passage relevance to improve retrieval accuracy in Japanese RAG pipelines. It is the second-smallest model in the v2 series, balancing high performance with extremely fast inference.

Architecture and strengths

Built on ModernBert (ruri-v3-pt-30m), the model uses 10 transformer layers with a hidden size of 256. Despite its compact size, it achieves an average benchmark score of 0.8699 across four Japanese retrieval datasets—outperforming many larger cross‑encoder variants from the v1 series. The model runs efficiently on GPU (6.5 s for ~150 k pairs using flash‑attention 2), on Apple Silicon (MPS), and on CPU. ONNX quantised versions (for AVX2 and ARM64) are also available for edge or CPU-only deployments, making the model suitable for low‑latency and resource‑constrained environments.

Benchmark performance

The model was trained using knowledge distillation from larger teacher models (japanese-reranker-cross-encoder-large‑v1 and japanese-bge-reranker-v2‑m3‑v1) and a single‑epoch training strategy to prevent overfitting. Training data includes multiple Japanese retrieval datasets plus hard negatives (hotchpotch/japanese-reranker-v2-hard-negatives).