qwen4b-mapper-50-50-regen-sft

DeepShopper Mapper (Qwen3-4B)

Stage 1 of the DeepShopper 2-step recommender brain (Mapper → retrieval → Reducer). The Mapper maps a user need → a structured, catalog-agnostic query plan: {num_items, reasoning, outfit:[{role, query}]} — it defines what to look for.

• Base: Qwen/Qwen3-4B, full SFT (3 epochs, lr 2e-5, eff-batch 8, max-len 2048).
• Data: flavianv/fashionrec-amz-mapper-50-50-sft — 10k gender-balanced rows; reasoning + per-slot queries generated in a single Qwen3-32B call per outfit (replaces an earlier procedural template).
• Eval (k=1 plan→retrieval→reward harness): mean reward 0.099 vs 0.021 zero-shot; pass@1>0.10 17%→55%, >0.15 3%→24%.
• Usage: system prompt = "DeepShopper Mapper… return JSON {num_items, reasoning, outfit:[{role, query}]}"; control item count via the need (e.g. "as a 4-item set" / "including top, bottom, footwear").
• Caveat: men's-side outputs can show a 2-item bias / role-label slips under free generation (AMZ-male source data).

Code: https://github.com/clijo/reco-rl (branch outfit_bundle).