Musical Instruments combined-data ranking head
Selected head-only checkpoint990 (epoch1.25), trained on3169 queries /12659 Bradley–Terry pairs. This repository contains the2560-weight scalar head, not a standalone backbone. Use the pinned SFT backbone in backbone.json and load score.pt into model.score after converting the head to float32. Backbone was frozen in BF16. Request + product titles are the only inputs; use the accompanying inference script for exact serialization.
Validation selection247/274=90.145985%, versus original236/274=86.131387%; final243/274=88.686131%. Checkpoint990 has both highest selection and lowest loss among eight checkpoints. This is validation-selected, not independent test performance. Wrong-query negatives contribute to pairwise loss but not same-query top-one selection. Ties fail. Synthetic partial-overlap negatives retained; title aliases can make ID labels ambiguous.
Training dataset. The507agent-dev holdout and275test were excluded and not evaluated for this selection. The274validation cohort has prior model-selection exposure. No claim of statistical significance.
Started from original head822 on SFT420; AdamW1e-4,batch4,accum4,seed42,twoepochs1584updates,every198evaluation. Selected scores are in step-990_eval.json. Final-head reload verification was exact; selected score.pt transfer is SHA256 verified. All original checkpoints are retained.