Product Matching - all-MiniLM-L6-v2

This is a specialized Sentence Transformer model fine-tuned for Product Matching and E-commerce Similarity tasks. It is based on all-MiniLM-L6-v2 and has been optimized to handle complex product titles, specifications, and search queries.

Core Capabilities

• Product Deduplication: Identify identical products across different listings with varying title formats.
• Semantic Search: Match search queries to product titles even when keywords don't match exactly.
• Spec Matching: Associate product titles with their corresponding technical specifications or descriptions.
• Cross-Category Retrieval: Trained across 25 distinct consumer and industrial categories.
• False Positive Removal: Specifically optimized to distinguish between similar but non-matching items (e.g., different product versions or models).

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: sentence-transformers/all-MiniLM-L6-v2
• Maximum Sequence Length: 256 tokens
• Output Dimensionality: 384 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset: 10 Million Synthetic Product Pairs (including Title-Title, Query-Title, and Title-Specs)

Training Focus: Reducing False Positives

Unlike generic embedding models, this model is trained to be stricter. It is specifically designed to:

• Distinguish between product versions (e.g., Sony XM4 vs XM5).
• Reject same-category cross-brand matches (e.g., Nike vs Adidas).
• Provide low similarity scores for unrelated products, ensuring higher precision in automated deduplication pipelines.

Training Domain

The model was fine-tuned on a diverse corpus covering:

• Consumer Goods: Electronics, Office Supplies, Furniture, etc.
• Industrial & B2B: Mechanical parts, Industrial equipment, Professional tools.

Model Sources

• Repository: surazbhandari/all-MiniLM-L6-v2-ProductMatching
• Documentation: Sentence Transformers Documentation

Full Model Architecture

SentenceTransformer(
 (0): Transformer({'max_seq_length': 256, 'do_lower_case': False, 'architecture': 'BertModel'})
 (1): Pooling({'word_embedding_dimension': 384, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
 (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference:

from sentence_transformers import SentenceTransformer
import util # Optional for cosine_sim if not using model.similarity

# 1. Load the model
model = SentenceTransformer("surazbhandari/all-MiniLM-L6-v2-ProductMatching")

# 2. Define product pairs
product_a = "Apple iPhone 15 Pro Max 256GB Titanium"
product_b = "iPhone 15 Pro Max - Blue Titanium - 256 GB"
unrelated_product = "Logitech MX Master 3S Wireless Mouse"

# 3. Encode product titles
embeddings = model.encode([product_a, product_b, unrelated_product])

# 4. Calculate Cosine Similarity
similarity_score = model.similarity(embeddings[0], embeddings[1])
print(f"Similarity (Product A & B): {similarity_score.item():.4f}")
# Expected: ~0.60 (Strong match for varied titles)

diff_similarity = model.similarity(embeddings[0], embeddings[2])
print(f"Similarity (Unrelated): {diff_similarity.item():.4f}")
# Expected: < 0.10 (Clear distinction)

Benchmarking

To evaluate the model's effectiveness, we compared it against the original all-MiniLM-L6-v2 model across various product matching scenarios.

Comparison Results

Summary Statistics

• Avg Negative Similarity: Fine-Tuned: 0.5132 | Base: 0.6198
• Avg Match Similarity: Fine-Tuned: 0.7687 | Base: 0.8759

Key Findings

While the base model often provides higher scores for direct matches, it is also much more "generous" with similar but different products. This fine-tuned model is significantly more effective at rejecting false positives, demonstrating lower similarity scores for hard negatives (different versions) and cross-brand comparisons. This makes it ideal for high-precision deduplication tasks.

Training Overview

The model was fine-tuned using a high-performance contrastive learning approach (MultipleNegativesRankingLoss).

Key Training Highlights:

• Dataset: 10 Million product-centric pairs.
• Optimization: Focused on maximizing the embedding distance between similar but non-identical products (Hard Negatives).
• Efficiency: Leveraged the lightweight all-MiniLM-L6-v2 architecture, maintaining fast inference speeds while significantly improving domain-specific accuracy.
• Convergence: Achievement of high precision within a single epoch of large-scale synthetic data training.

For full technical details on the base architecture, refer to the original all-MiniLM-L6-v2 model card.