oro-ai/qwen3-4b-shoppingbench-sft
The oro-ai/qwen3-4b-shoppingbench-sft is a 4 billion parameter Qwen3-based supervised fine-tuned language model developed by ORO-AI. It is specifically optimized for shopping agent tasks, achieving a 42.7% ASR on a leak-cluster-guarded held-out partition of the ShoppingBench SN15 corpus. This model is designed to distill a shopping agent from ShoppingBench subnet traces, making it suitable for e-commerce automation and agentic applications.
Loading preview...
Model Overview
The oro-ai/qwen3-4b-shoppingbench-sft is a 4-billion parameter language model based on the Qwen3 architecture, developed by ORO-AI. It has undergone supervised fine-tuning (SFT) using the leak-cluster-guarded ShoppingBench SN15 corpus. This model serves as a companion artifact for the research paper "Bittensor Agent Arenas as a Trajectory Primitive: Distilling a Shopping Agent from ShoppingBench Subnet Traces."
Key Capabilities
- Enhanced Shopping Agent Performance: Significantly improves upon the base Qwen3-4B model's performance on ShoppingBench, achieving a 42.7% Agent Success Rate (ASR) on a production-strict, held-out partition, compared to the base model's 18.0% ASR.
- Trajectory Primitive Distillation: Designed to distill a shopping agent from complex ShoppingBench subnet traces, enabling more effective automated shopping interactions.
- Ready-to-Use Integration: Provided as a merged full model, allowing direct loading with
transformersor serving with vLLM without the need for adapter stacking.
Training Data
The model was fine-tuned on a filtered corpus, oro-ai/sn15-shoppingbench-sft-15k, derived from raw traces available at oro-ai/sn15-shoppingbench-traces-18k.
Good For
- Developing and deploying automated shopping agents.
- Research into agentic behavior and trajectory primitives in e-commerce.
- Applications requiring specialized language understanding and generation for online shopping scenarios.