snapcart-ai/sam-1-base
SAM-1-Base is a 7.62 billion parameter language model developed by SnapCart AI, fine-tuned from Qwen2.5-7B-Instruct with a 32,768 token context length. It is specifically optimized for commerce reasoning tasks, achieving a 90.55 score on the SAM-Bench shopping assistant benchmark. This model excels at tasks like query understanding, product comparison, and review synthesis, making it suitable for e-commerce AI applications.
Loading preview...
SAM-1-Base: Specialized for E-commerce AI
SAM-1-Base (Shopping Agent Model) is a 7.62 billion parameter language model developed by SnapCart AI, built upon the Qwen2.5-7B-Instruct architecture. This model has been fine-tuned using LoRA adaptation, which is fully merged into the base weights, making it ready for direct inference. It features a substantial context length of 32,768 tokens and uses the ChatML format.
Key Capabilities & Performance
SAM-1-Base is specifically designed for 8 core shopping assistant capabilities, demonstrating strong performance on the proprietary SAM-Bench benchmark, where it scored 90.55/100. Its strengths include:
- Query Understanding: Achieves 98.37% in parsing shopping queries.
- Attribute Extraction: Scores 97.57% in extracting structured product attributes.
- Product Comparison: Reaches 94.88% in comparing products and selecting the best option.
- Purchase Decision: Scores 94.11% in making purchase recommendations.
- Review Synthesis: Achieves 92.45% in summarizing reviews with sentiment analysis.
While excelling in many areas, the model's performance in Product Recommendation (77.76%) and Personalization (77.59%) indicates areas for future improvement in preference modeling.
Intended Use Cases
- Research and Evaluation: Ideal for studying and comparing shopping assistant models.
- Prototyping: Suitable for developing e-commerce AI features.
- Academic Study: Useful for research into domain-specific LLM fine-tuning.
Limitations
It's important to note that SAM-1-Base was evaluated on synthetic data, is English-only, and lacks multimodal capabilities. It also inherits the biases and knowledge cutoff of its base model, Qwen2.5-7B-Instruct.