musical-instruments
musical-instrumentsmusical-instruments-id-overlap-reward-20260923-v1
Musical Instruments ID-overlap reward supervision, 2026-09-23 v1
14,000 positive rows (two randomly sampled distinct title orderings for each of 7,000 queries) and 22,513 usable unique negative rows: 36,513 labeled training rows. Positive selection uses Python Random seed 20260923, sorted query IDs, sampling two of four existing permutations without replacement. All original rows and selected indices are preserved. No generated negatives are replicated or replenished to force… See the full description on the dataset page: https://huggingface.co/datasets/flavianv/musical-instruments-id-overlap-reward-20260923-v1.musical-instruments-sound-datasetFTT_Musical_Instruments_trainmusical-instruments-sound-datasetThis is a processed dataset from this Kaggle dataset.
I've processed to hugging face format, added a validation split and, also fixed a couple of wrong labels that were present on the original dataset.
musical-instruments-ranker-unseen-20260923
Musical Instruments ranker data unseen by selected SFT checkpoint
All 2,217 positive examples not consumed by SFT checkpoint 420 are preserved in unseen_positives.jsonl. Twenty rows repeat a positive product set already seen by that checkpoint and are quarantined, leaving 2,197 eligible positives. These are exact query/set exclusions, not product-level or user-level independence.
Training/validation/test are grouped roughly 75%/12.5%/12.5%; eval means validation. Each positive… See the full description on the dataset page: https://huggingface.co/datasets/flavianv/musical-instruments-ranker-unseen-20260923.