Amazon ESCI (2M query-product pairs with relevance labels)
WANDS / Wayfair (231K query-product pairs)
GoShops (410K query-product pairs from real e-commerce search logs, click + purchase signals)

Total: 2.65M training pairs balanced across sources (150K per source).

Training Details

Base model: nvidia/llama-nemotron-rerank-1b-v2 (LLaMA 3.2 1B, bidirectional cross-encoder)
Loss: BinaryCrossEntropyLoss (pos_weight=3.0)
Epochs: 1 (optimal based on eval loss curve)
Batch size: 32
Learning rate: 2e-5 with 10% warmup
GPU: NVIDIA A100 80GB
Best eval loss: 0.635

Evaluation (NDCG@5)

Model	NDCG@5	Win/Tie/Loss
Base (nvidia)	0.9700	-
Fine-tuned	0.9957	7/2/1

+2.65% improvement over the base model on 10 e-commerce test queries (ES + EN).

Usage

python

from sentence_transformers.cross_encoder import CrossEncoder

model = CrossEncoder("scotto2/nemotron-ecommerce-reranker-1b", trust_remote_code=True)

pairs = [
    ("zapatillas running", "Zapatillas Running Nike Air Zoom"),
    ("zapatillas running", "Sarten antiadherente 28cm"),
]
scores = model.predict(pairs)
# [0.957, 0.044]