0xMeet/LLM-Personalized-Recommendation-Dataset

LLM Personalized Product Recommendation Dataset

A dataset prepared for an LLM-powered personalized product recommendation system combining user interaction history, natural-language shopping intent, product metadata, semantic retrieval, and personalized ranking.

Dataset

This dataset contains two configurations:

interactions

72,909 user-level examples

Split	Examples
Train	58,327
Validation	7,291
Test	7,291
Total	72,909

Each record contains:

{
  "user_id": "USER_ID",
  "history": [
    {
      "parent_asin": "PRODUCT_ID",
      "rating": 5.0,
      "timestamp": 1457328538000
    }
  ],
  "query": "wireless headphones for travel",
  "target": {
    "parent_asin": "PRODUCT_ID"
  }
}

The target contains only parent_asin. Rating and timestamp are intentionally excluded.

metadata

The complete preprocessed product catalog containing products from Amazon Fashion and Electronics.

Each product contains:

• main_category
• title
• average_rating
• rating_number
• features
• description
• price
• details
• parent_asin

features and description were limited to a maximum of five entries during preprocessing.

The complete catalog is retained independently of the interaction dataset so it can be used as a larger candidate pool for semantic retrieval and vector-store construction.

Query Generation

The natural-language queries are synthetic.

They were generated using Qwen2.5-3B-Instruct from product metadata and recent user interaction context.

The goal was to generate short shopping-intent queries rather than copy product titles or descriptions.

These queries should not be considered authentic Amazon search-log queries.

How to Use

Install the required library:

pip install -U datasets

Load interaction data

from datasets import load_dataset

DATASET_ID = "0xMeet/LLM-Personalized-Recommendation-Dataset"

interactions = load_dataset(
    DATASET_ID,
    "interactions"
)

train = interactions["train"]
validation = interactions["validation"]
test = interactions["test"]

print("Train:", len(train))
print("Validation:", len(validation))
print("Test:", len(test))

Load the complete metadata catalog

metadata = load_dataset(
    DATASET_ID,
    "metadata"
)

products = metadata["train"]

print("Products:", len(products))

Data Source

The underlying data are derived from Amazon Reviews 2023 by McAuley Lab.

Source paper:

Bridging Language and Items for Retrieval and Recommendation

Citation

@article{hou2024bridging,
  title={Bridging Language and Items for Retrieval and Recommendation},
  author={Hou, Yupeng and Li, Jiacheng and He, Zhankui and Yan, An and Chen, Xiusi and McAuley, Julian},
  journal={arXiv preprint arXiv:2403.03952},
  year={2024}
}

Query Generation Model

Synthetic queries were generated using Qwen2.5-3B-Instruct.

Source paper:

Qwen2.5 Technical Report

Citation

@article{qwen2.5,
  title={Qwen2.5 Technical Report},
  author={Yang, An and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and others},
  journal={arXiv preprint arXiv:2412.15115},
  year={2024}
}

Limitations

The natural-language queries are synthetic and may not fully represent real customer search intent.

Because the queries were generated using information associated with the target product, they should be treated as a synthetic recommendation benchmark, not as real search behavior.

Project

GitHub Repository