Abner0803/Amazon-Beauty-S1

This dataset is derived from Amazon Reviews'23 [1] Beauty category. The split is standard leave-one-out: the last item is the test target, the second-to-last is the validation target, and everything before that is training. The training portion is expanded by sliding window — every prefix becomes one example — so a user with a sequence of length $L$ contributes $L-3$ training rows with histories of length $1 \dots L-3$, one validation row and one test row.

Statistics

Users (= sequences)	9,880
Items	11,707
Interactions	81,767
Mean / median sequence length	8.28 / 6
Shortest / longest sequence	5 / 144
Rows: train / valid / test	52,127 / 9,880 / 9,880

Sequences are 5-core, so every user has at least five interactions and the shortest test history is four items. The test split touches all 11,707 items; the training split touches 11,131, so 576 items are never seen during training.

Files

File	Contents
train.jsonl	{"history": [asin, ...], "target": asin}
val.jsonl	same
test.jsonl	same
item_map.jsonl	full catalogue: adds description, category

1. Hou, Yupeng, et al. "Bridging language and items for retrieval and recommendation: Benchmarking LLMs as semantic encoders." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.