Sentence Transformers v4.0 Adds Reranker Training Support

The core of the Sentence Transformers v4.0 update is the introduction of a new training approach for Reranker (Cross-Encoder) models. Since a Reranker processes a query and document pair together to evaluate relevance, it offers more sophisticated retrieval performance than the existing Bi-Encoder (embedding) approach.

This guide covers in detail the key components for building a Reranker model:

• Dataset: Using local data (CSV, JSON, Parquet, etc.) and improving data quality through Hard Negatives Mining
• Training Configuration: Choosing an appropriate Loss Function, Training Arguments, and how to use the Trainer class
• Evaluation: Verifying model performance through various Evaluators

In practice, a ModernBERT-based Reranker model trained with this approach recorded performance surpassing existing large-scale general-purpose Reranker models, demonstrating that it's possible to directly build a powerful model optimized for specific domain data.