Cross-Encoder Training

When to Fine-Tune

Off-the-shelf rerankers (Cohere Rerank v3.5, BGE-reranker-v2-m3, Voyage rerank-2) are strong. Fine-tune when:

• Labeled gold set shows NDCG@10 plateauing around the vendor baseline.
• Your domain vocabulary drifts from web text (legal contracts, clinical notes, code).
• You must self-host a small, fast reranker (100-300 MB) for edge deployment.
• You need to distill a proprietary vendor model into something you own.

Skip fine-tuning if you have fewer than 2k labeled triples. Spend effort on better retrieval instead.

Training Data Shapes

Pointwise

(query, doc, label) where label in {0, 1} or a graded relevance score.

Pairwise