flavianv/musical-instruments-ranker-unseen-resampled-20260924-v1

Musical Instruments original ranker protocol: resampled v1

A training-only expansion of the original ranking dataset. Same 1,648 training queries, same original negative-generation rules. Four distinct orderings of each complete reference bundle give 6,592 positive rows. Two independent negatives per strategy are requested: random set, role collision, wrong item, wrong query. There are 13,150 negatives/comparisons (exactly twice the original 6,575): random set 3,296; role collision 3,262; wrong item 3,296; wrong query 3,296. Seventeen queries cannot support role collision under the original conservative rules; their 34 missing negatives are documented, not fabricated.

Each negative is paired once, cycling four positive orderings. This is not a Cartesian product. It increases orderings and sampled negatives, not independent queries. Partial-overlap negatives are deliberately retained. Wrong-query negatives retain the bundle but use an unrelated request; they contribute to pairwise loss, not same-query bundle selection. Labels are synthetic presumed preferences, not exhaustive relevance judgments.

Original validation (274 queries/1,092 pairs) and test (275/1,098) are copied byte-for-byte. No resampling or test selection. These are not the later 300-query SFT-generated pools. Query and reference-ID-set separation checked across all three splits. Products themselves can overlap. Queries were outside the selected SFT420 gradient prefix; later discarded SFT checkpoints consumed some. Title-only scoring cannot distinguish every ID-based label.

Files: train.jsonl paired training; positives.jsonl four orders; negatives.jsonl generated pair records; queries.jsonl source queries; eval.jsonl and test.jsonl original heldouts; manifest.json counts, seeds, deficits, provenance and checksums. No model weights. Seed 20260924 with increment 1000 for additional generation rounds.