KTO preference refinement (v3) applied on top of the merged opd_renderers SFT champion. Reaches 42.7% ASR on the 75-problem leak-cluster-guarded held-out partition (production-strict, temp 0), matching the SFT champion.

Companion artifact for the paper Bittensor Agent Arenas as a Trajectory Primitive: Distilling a Shopping Agent from ShoppingBench Subnet Traces. The published Qwen3-4B base scores 18.0% ASR on ShoppingBench; the distilled SFT-family models in this collection lift that to 42.7% on a leak-cluster-guarded held-out partition scored production-strict.

This is a merged full model (Qwen3-4B weights with the trained delta merged in), ready to load directly with transformers or serve with vLLM. No adapter stacking required.

Training data

Filtered corpus: oro-ai/sn15-shoppingbench-sft-15k
Raw traces: oro-ai/sn15-shoppingbench-traces-18k

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("oro-ai/qwen3-4b-shoppingbench-kto")
model = AutoModelForCausalLM.from_pretrained("oro-ai/qwen3-4b-shoppingbench-kto", torch_dtype="bfloat16", device_map="auto")

License

Apache-2.0, inherited from the Qwen3-4B base model.