⚡ ArcticDense: Jina Reranker v2 (InfoNCE Contrastive Champion)

This model is a fine-tuned Cross-Encoder reranker based on jinaai/jina-reranker-v2-base-multilingual (278M parameters) trained with Listwise InfoNCE Contrastive Cross-Entropy over actively mined hard negatives.

📊 Two-Stage Cascade Retrieval Performance

Evaluated on 1,000 queries searching across the full 64,552 Goodreads book catalog:

Pipeline Stage	Recall@1	Recall@5	Recall@10	MRR@10	Latency
Stage 1: Arctic Bi-Encoder Standalone	49.00%	66.40%	71.80%	0.5681	11.5 ms
Stage 2: Contrastive Hybrid Cascade (0.30 Dense + 0.70 Reranker)	52.50%	70.50%	75.00%	0.6001	182.8 ms
Improvement	+3.50%	+4.10%	+3.20%	+0.0320	—

🚀 Usage with Sentence-Transformers

from sentence_transformers import CrossEncoder

model = CrossEncoder("nansari0134/jina-reranker-v2-goodreads", max_length=256, trust_remote_code=True)

query = "cyberpunk detective investigation in dystopian megacity"
documents = [
    "Title: Neuromancer | Genres: sci-fi, cyberpunk | Description: Case was the sharpest data-thief in the matrix...",
    "Title: Pride and Prejudice | Genres: romance, classics | Description: A story of manners, marriage, and society..."
]

pairs = [[query, doc] for doc in documents]
scores = model.predict(pairs)
print("Reranked Scores:", scores)

Loading…