DeepShopper Reward V0 — AMZ bundle likelihood judge (Qwen3-4B)
Reward V0 for DeepShopper: a Qwen3-4B SFT'd to generate the matching outfit bundle
for a natural-language need (P(outfit | need)). Used as a reward via teacher-forced
likelihood of a candidate outfit (100·exp(−mean NLL)), with hard gates (invalid /
duplicate / role-category).
- Data:
flavianv/apparel23-qwen32b-kept-outfits-with-products(AMZ Apparel'23 gold/kept bundles). - Role: in-distribution relative signal (SFT mapper ≫ zero-shot).
- Known limits: likelihood ≠ quality; gender-blind and reasoning-blind; in-distribution (AMZ) only. Superseded for discrimination by the pairwise V1 (
flavianv/qwen4b-reward-pairwise-v1).
Code: https://github.com/clijo/reco-rl (branch outfit_bundle).