DeepShopper Reducer V0 (Qwen3-4B)

• Base: Qwen/Qwen3-4B, full SFT (1 epoch, 8k-row subset, max-len 2048).
• Data: flavianv/deepshopper-reducer-sft — need + plan + per-role candidate pool (+~30% distractors) → pick the right candidate by ephemeral handle (anti-memorization). Target = soft-gold (closest retrievable to the purchased item, gender-checked).
• Held-out results: distractor-selection 1.4% vs 33% random; end-to-end grid: distractor 0.3–1.9% vs ~21% top-1, gender-mismatch 1.7–4.6% vs 17.7% (filter off) — i.e. it internalizes the gender filter. Valid/count 1.0; generalizes k=5→k=10.
• Scope: proven as a guard/selector (reject wrong-gender/role/mislabeled, structure); fine intra-slot ranking is low-headroom (next: rating×log-reviews tie-break). No reasoning yet (V0 = direct selection).