Training small language models on retail catalog data

Training small language models on retail catalog data involves fine-tuning lightweight architectures—typically under 10 billion parameters—on structured and unstructured product information to achieve high precision in domain-specific tasks. This process enables a model to understand niche product attributes, technical specifications, and inventory relationships more accurately than a generic large language model (LLM). For mid-size retailers, this approach reduces latency and operational costs while keeping sensitive catalog data within a private environment.

While general-purpose models like GPT-4 are highly capable, they often struggle with the specific nomenclature of specialized industries, such as industrial hardware, high-end electronics, or niche apparel. A custom-trained Small Language Model (SLM) bridges this gap by internalizing the specific patterns of your SKU library, making it an essential tool for custom SLM models designed for internal search, automated tagging, and customer support agents.

Why SLMs Outperform General Models for Retail Catalogs

Retail catalogs are dense with technical specifications that general models often misinterpret or hallucinate. For example, a general model might not distinguish between a "1/4-20 x 1-inch Grade 5 zinc-plated hex bolt" and a similar item with a different thread pitch or grade.

Small Language Models (SLMs) such as Microsoft’s Phi-3, Meta’s Llama 3 8B, or Mistral 7B offer several advantages for retail-specific applications:

1. Latency: SLMs can generate responses in milliseconds, which is critical for real-time product discovery and search suggestions.
2. Context Efficiency: Because the model is pre-trained or fine-tuned on your specific catalog, you do not need to provide massive product descriptions in every prompt, saving on "context window" costs.
3. Privacy: These models can run on single-node GPU instances or even on-premise hardware, ensuring your inventory strategy and pricing data never leave your infrastructure.
4. Accuracy on Niche Attributes: Fine-tuning allows the model to learn the "language" of your brand, including proprietary sizing, material names, and compatibility rules.

Preparing Your Retail Catalog for Model Training

The quality of your SLM is directly proportional to the cleanliness of your training data. Most retail data resides in ERP (Enterprise Resource Planning) or PIM (Product Information Management) systems in a format that is too fragmented for direct training.

Data Cleaning Checklist

Before training begins, your team must perform a data audit. Use the following checklist to ensure the catalog is ready:

• Deduplication: Remove redundant entries for the same product across different regions or warehouses.
• Attribute Normalization: Ensure units of measurement are consistent (e.g., all weights in grams or ounces, not a mix of both).
• Image-to-Text Enrichment: If your catalog has sparse descriptions, use a vision-language model to generate text descriptions from product images to expand the training corpus.
• Hierarchy Mapping: Clearly define the relationship between categories, sub-categories, and individual SKUs.

Formatting the Data

For fine-tuning, data is typically converted into a JSONL (JSON Lines) format. Each line should represent a "conversation" or a "completion" task. For a retail catalog, you should create pairs that reflect how the model will be used.

Example JSONL Training Pair:

{"instruction": "Identify the material and thread count for SKU-9928.", "input": "Product: Egyptian Cotton Sheet Set. Specs: 800TC, 100% long-staple cotton, Sateen weave.", "output": "The material is 100% long-staple Egyptian cotton with a thread count of 800."}

Selecting a Base Architecture

Not all small models are created equal. The choice of base model depends on your specific use case, whether it is product discovery with SLM or automated inventory tagging.

For most retail applications focusing on technical specs, the Llama 3 8B or Mistral 7B architectures provide the best balance between reasoning capability and speed.

The Step-by-Step Process of Training Small Language Models on Retail Catalog Data

Once the data is prepared and the model is selected, the engineering team follows a structured pipeline to fine-tune the model for inventory and discovery tasks.

Step 1: Supervised Fine-Tuning (SFT)

In this phase, the model is fed thousands of examples of your catalog data. We use a technique called QLoRA (Quantized Low-Rank Adaptation). This method significantly reduces the memory requirements for training by only updating a small subset of the model's weights while keeping the rest frozen. This is the most cost-effective way for SMEs to achieve high-performance results.

Step 2: Domain Adaptation

If your retail niche uses highly specialized language (e.g., medical supplies or aerospace components), we perform domain adaptation. This involves feeding the model raw technical manuals, whitepapers, and full-length catalog descriptions without specific instructions. This helps the model understand the "grammar" of your specific industry.

Step 3: Instruction Tuning

This is where we teach the model how to behave. If the goal is a retail assistant, we tune it to be helpful and concise. If the goal is data cleaning, we tune it to output strictly in JSON or CSV formats. For retail, this often involves "negative training"—teaching the model to say "I do not know" when a product is out of stock rather than hallucinating a replacement.

Step 4: Evaluation via RAG Benchmarking

Before deployment, we test the SLM against a Retrieval-Augmented Generation (RAG) system. We ask the model complex questions about the inventory and measure its accuracy against a known "ground truth" dataset.

Example Evaluation Metric:

• Attribute Accuracy: Does the model correctly identify the 'Voltage' for 100% of the power tools in the test set?
• Constraint Adherence: Does the model respect 'out of stock' flags when suggesting products?

Common Mistakes in Retail SLM Implementation

In our experience at ZEON, we see three recurring errors when companies attempt to build custom SLMs for ecommerce:

1. Over-fitting on Outdated Inventory: If you train a model too rigidly on current stock, it may struggle when the catalog refreshes next season. We recommend training on product types and attributes rather than just specific SKU IDs.
2. Neglecting the "Cold Start" Problem: New products added after the model is trained will not be "known" by the model's internal weights. A hybrid approach—using the SLM to understand the query and a vector database (RAG) to fetch the latest inventory—is often the most robust solution.
3. Dirty SKU Logic: Many catalogs use inconsistent SKU logic (e.g., using both "-" and "_" as delimiters). Without normalization, the model will treat these as different entities, leading to fragmented search results.

When is a Custom SLM Not Worth the Investment?

While powerful, training small language models on retail catalog data is not always the right move. It may be overkill if:

• Your catalog is small: If you have fewer than 500 products, a well-engineered prompt using a standard LLM (like GPT-4o) with a simple RAG system will likely suffice.
• Your data is in total disarray: If your PIM is not centralized and you cannot produce a clean CSV of your products, the model will simply learn your mistakes.
• Low Query Volume: If your internal teams or customers only perform a few dozen searches a day, the cost of hosting a dedicated model may exceed the benefits of lower latency.

Implementation Roadmap for This Week

If you are ready to move beyond the hype and build a functional model, here is how to start:

1. Days 1-2: Data Extraction. Export your top 5,000 SKUs with all available metadata (descriptions, specs, categories) into a single flat file.
2. Day 3: Clean and Format. Use a script to normalize units and remove HTML tags from descriptions. Convert the data into the JSONL format mentioned above.
3. Day 4: Baseline Testing. Use an open-source tool like Ollama to run a local instance of Llama 3 or Phi-3. Ask it 10 technical questions about your products. Note where it fails.
4. Day 5: Feasibility Review. Evaluate if the failures are due to a lack of knowledge (which fine-tuning fixes) or a lack of current data (which RAG fixes).

For companies managing complex inventories, the transition from generic AI to specialized SLMs represents a shift toward higher operational efficiency. By teaching a model the specifics of your catalog, you ensure that your AI tools are as knowledgeable as your best floor manager or lead engineer.