ShoppingBench SN15: Trajectory Primitive - a oro-ai Collection

oro-ai 's Collections

ShoppingBench SN15: Trajectory Primitive

ShoppingBench SN15: Trajectory Primitive

updated Jun 10

Datasets + distilled Qwen3-4B models from the ORO Subnet 15 trajectory-primitive paper (raw/filtered corpora, SFT, rejection, SFT+KTO).

oro-ai/sn15-shoppingbench-traces-18k

Preview • Updated Jun 10 • 86

Note Raw 18,043 SN15 race traces (pre-filter SFT input)
oro-ai/sn15-shoppingbench-sft-15k

Preview • Updated Jun 10 • 124

Note Leak-cluster-guarded 15,269 train + 2,676 eval holdout
oro-ai/qwen3-4b-shoppingbench-sft

Text Generation • 4B • Updated Jun 10 • 22

Note Base SFT stage
oro-ai/qwen3-4b-shoppingbench-rejection

Text Generation • 4B • Updated Jun 10 • 18

Note Reward-weighted rejection-sampled stage
oro-ai/qwen3-4b-shoppingbench-kto

Text Generation • 4B • Updated Jun 10 • 23

Note SFT + KTO preference refinement (42.7% ASR)
Bittensor Agent Arenas as a Trajectory Primitive: Distilling a Shopping Agent from ShoppingBench Subnet Traces

Paper • 2606.10064 • Published Jun 8