Datasets:

---

hover-nlp

/

hover

like 7

Follow

hover-nlp 8

Summary

This PR adds Parquet-based data files to enable scriptless loading of the HoVer dataset, eliminating the need for trust_remote_code.

Changes

• ✅ Added data/train.parquet (18,171 examples)
• ✅ Added data/validation.parquet (4,000 examples)
• ✅ Added data/test.parquet (4,000 examples)
• ✅ Updated README.md with:
  • Quick start guide
  • Detailed data field descriptions
  • Corrected MIT license information
  • Complete citations and documentation

Benefits

1. No trust_remote_code required - Users can load the dataset with standard load_dataset("hover-nlp/hover") without security warnings
2. Faster loading - Parquet format is more efficient than JSON parsing
3. 100% compatible - Data structure is identical to the existing hover.py script output
4. Better CI/CD support - Modern pipelines prefer scriptless formats

Verification

The Parquet format has been verified to produce identical output to the original hover.py script:

• All field types match (ClassLabel for labels, etc.)
• Supporting facts structure preserved: [{"key": "title", "value": idx}, ...]
• All 26,171 examples converted successfully

Usage

from datasets import load_dataset

# No trust_remote_code needed!
dataset = load_dataset("hover-nlp/hover")

Backward Compatibility

The existing hover.py script remains in the repository for backward compatibility. Users can continue using the script-based approach if needed.

Testing

Dataset tested and confirmed working at: https://huggingface.co/datasets/vincentkoc/hover-parquet

Source branch: {source_repo}:{source_branch}