japanese-reranker-cross-encoder-small-v1

Japanese Reranker Cross-Encoder (Small)

hotchpotch/japanese-reranker-cross-encoder-small-v1 is a cross-encoder (reranker) model trained on Japanese text. It accepts a query and a passage and outputs a relevance score, making it suitable for re-ranking candidate lists in retrieval pipelines.

Key Strengths

• 12 transformer layers with a hidden size of 384 – a balance of performance and inference speed.
• Part of a family of Japanese rerankers (xsmall, small, base, large, and a BGE-based variant) allowing users to scale model size as needed.
• Trained on Japanese data; outperforms general multilingual rerankers across multiple Japanese benchmarks.

Best For

• Re-ranking search results or retrieval outputs in Japanese-language applications.
• Scenarios where a lightweight model reduces latency without sacrificing ranking quality.

Benchmark Results

The following table reports normalized discounted cumulative gain (nDCG@10) on standard Japanese IR datasets. For comparison, results from other popular rerankers and baselines are included.