yacud1/arxiv-bge-base-818k

arXiv Abstracts — BGE-base Embeddings + FAISS Index (818k papers)

Pre-built vector search index over 817,961 arXiv papers (titles + abstracts), embedded with BAAI/bge-base-en-v1.5 into 768-dim vectors and packaged as a ready-to-query FAISS index.

Why this exists: embedding ~818k abstracts through BGE-base is the slow, hardware-hungry step. Download this once and you skip it entirely — load the index, embed only your query string with the same model, and search. No re-vectorizing the corpus.

Contents

File	Size	Description
index.faiss	~2.3 GB	FAISS IVF-PQ index of all 817,961 vectors (768-dim, bge-base)
ids.npy	~50 MB	str array of arXiv IDs; row i of the index ↔ ids[i]
metadata.parquet	~341 MB	Per-paper metadata: arXiv ID, title, abstract, authors, categories, date
manifest.json	<1 KB	Build provenance: model, dim, doc count, timestamp

Alignment: index.faiss, ids.npy, and metadata.parquet are all in the same row order. FAISS returns integer positions → map through ids.npy → join to metadata.parquet.

Model & index details

• Embedding model: BAAI/bge-base-en-v1.5 (768-dim, cosine / inner-product)
• Text embedded: title + "\n" + abstract per paper
• Index type: FAISS IVF-PQ (compressed; approximate nearest neighbor)
• Documents: 817,961
• Built: 2026-05-28

Important — the index is model-locked. Query vectors must come from BAAI/bge-base-en-v1.5. Embedding your query with any other model returns meaningless results.

IVF-PQ is lossy. Vectors are product-quantized for compact fast search; you cannot recover the exact float32 embeddings from index.faiss. This is ideal for retrieval, not for re-ranking on raw vectors. (If you need raw float32 embeddings, open an issue / discussion.)

Coverage

This is a quantitative / ML / stats-focused slice of arXiv, not the full corpus. Computer science, mathematics, and statistics make up ~88% of papers, with strong quantitative-finance and economics representation. Hard-physics, astro, and high-energy categories are present only in small numbers.

Counts below are by primary category (each paper's first listed category); percentages are of all 817,961 papers.

Top-level group	Papers	Share
cs — Computer Science	475,899	58.2%
math — Mathematics	186,397	22.8%
stat — Statistics	61,860	7.6%
eess — Electrical Eng. & Systems Science	22,446	2.7%
physics — Physics (general/interdisc.)	20,053	2.5%
q-fin — Quantitative Finance	13,457	1.6%
q-bio — Quantitative Biology	7,911	1.0%
quant-ph — Quantum Physics	7,452	0.9%
econ — Economics	7,156	0.9%
cond-mat — Condensed Matter	5,278	0.6%
math-ph — Mathematical Physics	3,577	0.4%
others (astro-ph, nlin, gr-qc, hep-*, nucl-*, legacy codes)	~6,500	0.8%

Notable subcategories

Largest subcategories within the main groups (by primary category):

• cs — cs.LG Machine Learning (132,294) · cs.CL Computation & Language (80,653) · cs.CV Computer Vision (44,707) · cs.IT Information Theory (39,497) · cs.AI Artificial Intelligence (37,935) · cs.DS Data Structures & Algorithms (18,108) · cs.SI Social & Information Networks (11,750) · cs.RO Robotics (10,895) · cs.CR Cryptography & Security (10,509) · cs.GT Computer Science & Game Theory (9,263)
• math — math.PR Probability (43,874) · math.OC Optimization & Control (38,899) · math.NA Numerical Analysis (34,957) · math.DS Dynamical Systems (25,761) · math.ST Statistics Theory (16,615) · math.CO Combinatorics (7,322)
• stat — stat.ME Methodology (25,310) · stat.ML Machine Learning (20,605) · stat.AP Applications (10,929) · stat.CO Computation (4,231)
• q-fin — q-fin.ST Statistical Finance (2,325) · q-fin.MF Mathematical Finance (1,875) · q-fin.GN General Finance (1,709) · q-fin.RM Risk Management (1,569) · q-fin.CP Computational Finance (1,547) · q-fin.PM Portfolio Management (1,421) · q-fin.PR Pricing of Securities (1,323) · q-fin.TR Trading & Market Microstructure (1,305)
• econ — econ.GN General Economics (4,924) · econ.EM Econometrics (1,522)
• eess — eess.SP Signal Processing (6,763) · eess.IV Image & Video Processing (6,595) · eess.SY Systems & Control (5,346) · eess.AS Audio & Speech Processing (3,742)

Papers are frequently cross-listed. The full categories field for each paper (all cross-lists, space-separated) is preserved in metadata.parquet, so you can filter to any subcategory yourself.

Quick start

from huggingface_hub import snapshot_download
import faiss, numpy as np, pandas as pd
from sentence_transformers import SentenceTransformer

# 1. Download the index files
path = snapshot_download(
    repo_id="yacud1/arxiv-bge-base-818k",
    repo_type="dataset",
    allow_patterns=["index.faiss", "ids.npy", "metadata.parquet"],
)

# 2. Load index + alignment + metadata
index = faiss.read_index(f"{path}/index.faiss")
ids   = np.load(f"{path}/ids.npy", allow_pickle=True)
meta  = pd.read_parquet(f"{path}/metadata.parquet").set_index("id")

# 3. Embed ONLY your query with the SAME model
model = SentenceTransformer("BAAI/bge-base-en-v1.5")
q = model.encode(["optimal execution under market impact"],
                 normalize_embeddings=True).astype("float32")

# 4. Search
scores, pos = index.search(q, k=10)
for score, i in zip(scores[0], pos[0]):
    row = meta.loc[ids[i]]
    print(f"{score:.3f}  arXiv:{ids[i]}  {row['title']}")

BGE retrieval tip: for short queries, prepend the recommended instruction "Represent this sentence for searching relevant passages: " before encoding.

Loading just the metadata

from datasets import load_dataset
meta = load_dataset("yacud1/arxiv-bge-base-818k", data_files="metadata.parquet")

Provenance & licensing

• Metadata & abstracts originate from the arXiv metadata corpus, released by arXiv under CC0 1.0 (public domain). This dataset redistributes that metadata under the same terms. Thank you to arXiv for open-access interoperability. See https://arxiv.org and the arXiv Kaggle metadata dataset.
• Embeddings / FAISS index are a derived artifact produced with the open-source BAAI/bge-base-en-v1.5 model.
• This dataset contains no full-text PDFs — titles, abstracts, and metadata only.

Intended uses

• Semantic search / retrieval over arXiv abstracts
• RAG over scientific literature
• Nearest-neighbor / "related papers" discovery
• Bootstrapping a domain corpus without paying the embedding cost

Limitations

• Coverage is a snapshot as of the build date (2026-05-28); newer papers absent.
• ANN search is approximate — recall depends on nprobe. Raise index.nprobe (e.g. index.nprobe = 32) to trade speed for recall.
• English-language abstracts; non-English content is out of distribution for the model.

Citation

If you use arXiv metadata, please follow arXiv's usage guidance and cite the relevant papers directly. For the embedding model, cite the BGE / FlagEmbedding work by BAAI.