ismailemir/arxiv-indices

ArXiv Search Indices

Pre-built search indices for ArXiv paper corpus supporting multiple retrieval methods.

📊 Contents

File	Description	Format
bm25_index.zip	BM25 lexical search index	Compressed folder
tfidf_index.npz	TF-IDF term frequency index	NumPy compressed
scibert_embeddings.npy	SciBERT semantic embeddings	NumPy array

🚀 Quick Start

from huggingface_hub import hf_hub_download
import zipfile
import bm25s
import numpy as np

# Download and extract BM25
bm25_zip = hf_hub_download("ismailemir/arxiv-indices", "bm25_index.zip", repo_type="dataset")
with zipfile.ZipFile(bm25_zip, 'r') as zip_ref:
    zip_ref.extractall("./bm25_index")

bm25_retriever = bm25s.BM25.load("./bm25_index")

# Download TF-IDF
tfidf_path = hf_hub_download("ismailemir/arxiv-indices", "tfidf_index.npz", repo_type="dataset")
tfidf_data = np.load(tfidf_path, allow_pickle=True)

# Download SciBERT embeddings
scibert_path = hf_hub_download("ismailemir/arxiv-indices", "scibert_embeddings.npy", repo_type="dataset")
embeddings = np.load(scibert_path)

🔍 Retrieval Methods

BM25 (Lexical)

• Best for exact keyword matching
• Fast and efficient
• Great for technical terms

TF-IDF (Term Frequency)

• Statistical word importance
• Good for general search
• Lightweight and fast

SciBERT (Semantic)

• Deep learning embeddings
• Understands context and meaning
• Best for conceptual search

📚 Related

• 📖 Corpus: ismailemir/arxiv-corpus
• 🔬 Source: Cornell ArXiv Dataset
• 🤖 Model: SciBERT

📄 License

Apache 2.0 - Please cite ArXiv if using this data.

🙏 Citation

@article{clement2019arxiv,
  title={On the Use of ArXiv as a Dataset},
  author={Clement, Colin B and Bierbaum, Matthew and O'Keeffe, Kevin P and Alemi, Alexander A},
  journal={arXiv preprint arXiv:1905.00075},
  year={2019}
}