ramitha2002/genieai-product-reranker

GenieAI Gift Product Reranker

This CrossEncoder reranks ecommerce products for English gift-shopping queries. It scores each (query, product text) pair; a higher score means the product should rank higher.

Model Details

Item	Value
Repository	ramitha2002/genieai-product-reranker
Base model	cross-encoder/ms-marco-MiniLM-L6-v2
Architecture	MiniLM CrossEncoder
Language	English
Maximum input length	384 tokens
License	Apache 2.0

Training Data

The model was fine-tuned on the public tasksource/esci mirror of Amazon's Shopping Queries Dataset. Only English/US Task 1 query groups matching these gift categories were selected:

cakes and desserts
flower bouquets
chocolates and candy
perfume and fragrance
jewelry
fashion and accessories
gift baskets and hampers
skincare and beauty sets
personalized gifts
home decor and candles

All candidates belonging to each selected query were retained, including irrelevant negatives.

Split	Pairs
Train	20,486
Validation	8,677
Validation query groups	424

Training-category distribution:

Category	Pairs
Fashion and accessories	6,608
Jewelry	6,551
Flower bouquets	1,707
Chocolates and candy	1,602
Home decor and candles	1,328
Cakes and desserts	1,194
Perfume and fragrance	834
Gift baskets and hampers	240
Skincare and beauty sets	222
Personalized gifts	200

ESCI labels were converted to numeric relevance targets:

Exact       = 1.00
Substitute  = 0.70
Complement  = 0.35
Irrelevant  = 0.00

Training Settings

Setting	Value
Epochs	2
Train batch size	32
Evaluation batch size	64
Learning rate	2e-5
Loss	Binary cross-entropy
Train loss	0.5644
Final evaluation loss	0.5207

Evaluation

Evaluation used complete, grouped ESCI validation queries.

Metric	Result
NDCG@10	0.8796
MRR	0.9591
Hit Rate@4	0.9906

These results measure the filtered ESCI validation set and do not guarantee the same performance on GenieAI's live catalog.

Usage

from sentence_transformers import CrossEncoder

model = CrossEncoder("ramitha2002/genieai-product-reranker")

query = "birthday flowers for mother"
products = [
    "Title: Pink rose bouquet\nDescription: Fresh roses for birthdays",
    "Title: Wireless gaming mouse\nDescription: RGB computer mouse",
]

scores = model.predict([(query, product) for product in products])
ranked = sorted(zip(products, scores), key=lambda item: item[1], reverse=True)
print(ranked)

Raw outputs are ranking scores, not calibrated probabilities. Compare scores only among products evaluated for the same query.

Intended Use

• Rerank approximately 30 products retrieved by RAG or search.
• Apply hard stock, delivery, and budget filters before or after retrieval.
• Return the best four products after reranking.
• Fall back to the original retrieval order if model inference fails.

Limitations

• Optimized for English gift-product searches.
• Fashion and jewelry are overrepresented in the training set.
• Weaker performance is expected for skincare, personalized gifts, and hampers.
• The dataset does not represent GenieAI's live prices, stock, delivery rules, or complete catalog.
• Validate the model on real GenieAI queries before production use.

Data and Base Model

• Dataset: https://huggingface.co/datasets/tasksource/esci
• Original ESCI source: https://github.com/amazon-science/esci-data
• Base model: https://huggingface.co/cross-encoder/ms-marco-MiniLM-L6-v2