thulthula/cvpr-10k-390k

CVPR 2026 - Amazon Reviews Training Data

Training data for LoViF @ CVPR 2026 challenge.

Format

Parquet file (compressed) with columns:

• parent_asin: Product ID
• main_category: Category name
• title: Product title
• description: Product description
• features: Product features
• average_rating: Average rating (1-5 stars)

Data Quality

• ✅ All samples have complete data (no missing fields)
• ✅ setA eval products excluded (no data leakage)
• ✅ Only products with non-empty title, description, features, and rating

Usage

import pandas as pd

# Download and load (parquet is faster and smaller!)
df = pd.read_parquet("hf://datasets/YOUR_USERNAME/cvpr-2026-amazon-training/all_categories_training.parquet")

# Combine text features
df['text'] = df.apply(lambda row:
    f"Title: {row['title']}\n\n"
    f"Description: {row['description']}\n\n"
    f"Features: {row['features']}",
    axis=1
)

# Use for training
X = df['text']
y = df['average_rating']

Citation

If you use this dataset, please cite the original Amazon Reviews 2023 dataset:

@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}
}