oro-ai/sn15-shoppingbench-traces-18k
ShoppingBench SN15 Race Traces (18K, winners, unanimous-50) Paper: arXiv:2606.10064Code: https://github.com/ORO-AI/shoppingbench-trajectory-primitive Multi-turn agentic shopping trajectories harvested from ORO Subnet 15 (SN15), the Bittensor deployment of the ShoppingBench agentic-commerce benchmark. These are the unfiltered 18K traces referenced in the paper Bittensor Agent Arenas as a Trajectory Primitive: Distilling a Shopping Agent from ShoppingBench Subnet Traces. This… See the full description on the dataset page: https://huggingface.co/datasets/oro-ai/sn15-shoppingbench-traces-18k.
ShoppingBench SN15 Race Traces (18K, winners, unanimous-50)
Paper: arXiv:2606.10064 Code: https://github.com/ORO-AI/shoppingbench-trajectory-primitive
Multi-turn agentic shopping trajectories harvested from ORO Subnet 15 (SN15), the Bittensor deployment of the ShoppingBench agentic-commerce benchmark. These are the unfiltered 18K traces referenced in the paper Bittensor Agent Arenas as a Trajectory Primitive: Distilling a Shopping Agent from ShoppingBench Subnet Traces.
This corpus is the SFT input before the structural-quality / leak-cluster filter. For the filtered, leak-cluster-guarded training corpus with a held-out eval split, see the companion dataset oro-ai/sn15-shoppingbench-sft-15k.
Contents
Format
Each row is NeMo SFT OpenAI shape:
{"messages": [...], "tools": [...4 tool schemas...], "task_name": "trac_shoppingbench_sft", "trac_export_version": "trac-nemo-sft-openai-v0.1-oro-converted"}<think> reasoning content is stripped (keep_think=False) to avoid narration-collapse during SFT.
Provenance
Built from a 146K-trajectory race-phase extract of SN15 (92% race-bank coverage). The unanimous-50 filter keeps winning trajectories that the LLM reasoning judge scored consistently. See the paper for the arena mechanics (race system, reasoning judge, rotating leak-cluster-guarded problem suite).
License
CC BY 4.0. Please cite the paper if you use this corpus.