multilingual-e5-large-finetuned-orders

Fine-tuned version of intfloat/multilingual-e5-large for order-offer matching task.

Model Description

This model was fine-tuned on a Russian dataset of order-offer pairs for semantic similarity matching. It significantly outperforms the base model on this specific task.

Training Details

• Base model: intfloat/multilingual-e5-large
• Training data: 68,270 order-offer pairs
• Loss function: MultipleNegativesRankingLoss
• Epochs: 3
• Batch size: 32
• Learning rate: 2e-5
• Training time: ~22 minutes on NVIDIA RTX PRO 6000

Performance

Metric	Base E5	Fine-tuned	Improvement
Accuracy@1	49.40%	72.93%	+23.53%
Accuracy@5	69.52%	91.20%	+21.67%
Accuracy@10	76.83%	95.11%	+18.28%
MRR	0.586	0.811	+0.225

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('olegGerbylev/multilingual-e5-large-finetuned-orders')

# Important: Use E5-style prefixes
orders = ["query: Кабель ВВГнг 3x2.5 | 100 м"]
offers = ["passage: Кабель ВВГнг(А)-LS 3х2,5 | 100.0 м"]

order_embeddings = model.encode(orders)
offer_embeddings = model.encode(offers)

# Compute similarity
from sklearn.metrics.pairwise import cosine_similarity
similarity = cosine_similarity(order_embeddings, offer_embeddings)
print(f"Similarity: {similarity[0][0]:.4f}")

Input Format

This model expects E5-style prefixes:

• For queries (orders): "query: <text>"
• For documents (offers): "passage: <text>"

License

MIT