igriv/dire-arxiv-bge-small-embeddings

DiRe arXiv BGE-small paper-level embeddings

Paper-level mean-pooled embeddings for 723,457 arXiv papers, produced with BAAI/bge-small-en-v1.5, together with arXiv descriptive metadata, 2-d and 3-d DiRe / UMAP layouts of the full corpus, and provenance manifests.

This is the dataset underlying the arXiv experiment in:

Kolpakov, A. and Rivin, I. DiRe: Topology-Faithful Dimensionality Reduction. (PNAS submission, 2026.) Igor Rivin: ORCID 0000-0001-9302-2169.

The pipeline is: 130 M LaTeX-derived chunks → unit-normalized BGE-small chunk embeddings → mean-pool per paper (one 384-d vector per paper) → DiRe / UMAP projection. The released embeddings are the mean-pooled, un-renormalized vectors (norms ~0.83-1.0). Downstream code typically L2-renormalizes them before further use.

What's in the release

Config	File(s)	Rows	Schema
default	embeddings/part-*.parquet (8 shards, ~1.03 GB)	723,457	arxiv_id: string, embedding: fixed_size_list<float32, 384>
metadata	metadata.parquet	723,457	arxiv_id, title, primary_category, categories: list<string>, n_chunks
layouts_2d	layouts/layouts_2d.parquet	7,234,570	arxiv_id, method ∈ {dire, umap}, n_neighbors ∈ {8,16,32,64,128}, x, y
layouts_3d	layouts/layouts_3d.parquet	1,446,914	arxiv_id, method, n_neighbors=16, x, y, z

Provenance lives in provenance/:

• embedding_config.json — model, pooling, normalization, embedding dim
• software_versions.json — Python / library versions used at build time
• file_hashes.json — SHA-256 of every released file

Quick start

from datasets import load_dataset
import numpy as np

# 384-d paper embeddings (default config)
ds = load_dataset("igriv/dire-arxiv-bge-small-embeddings", split="train")
X = np.vstack(ds["embedding"]).astype("float32")   # (723457, 384)
ids = np.asarray(ds["arxiv_id"])

# arXiv descriptive metadata
meta = load_dataset("igriv/dire-arxiv-bge-small-embeddings",
                    name="metadata", split="train").to_pandas()

# Pre-computed 2-d layouts (filter to the canonical setting used in the paper)
layouts = load_dataset("igriv/dire-arxiv-bge-small-embeddings",
                       name="layouts_2d", split="train").to_pandas()
canonical = layouts[(layouts.method == "dire") & (layouts.n_neighbors == 16)]

A minimal loader is also included as load_example.py.

Embedding details

Model	BAAI/bge-small-en-v1.5
Embedding dim	384
Source text	LaTeX-derived chunks (paper body + abstract)
Number of chunks pooled	130,028,430 across 723,457 papers (~180/paper avg)
Pooling	mean over chunk vectors (each chunk vector is BGE's L2-unit output)
Released dtype	float32
Released normalization	un-normalized mean-pool (typical norm 0.83–1.0); re-normalize to L2 unit before use if you want cosine ↔ Euclidean equivalence
Per-paper chunk count	available as n_chunks in metadata

Layouts

The 2-d layouts are the full corpus projected down to 2-d for both DiRe (force-directed, GPU; dire-rapids) and cuML UMAP. The sweep over n_neighbors ∈ {8, 16, 32, 64, 128} is included so readers can reproduce every figure in the paper directly. The canonical setting is n_neighbors=16, method=dire. 3-d layouts are provided at n_neighbors=16 for both methods.

Layout coordinates are in raw method output units (arbitrary; not standardized).

Reproducing the embeddings

The pooled embeddings here are the only thing required to reproduce the paper's arXiv experiments (DiRe / UMAP layouts, kNN-preservation evaluation, Betti-curve DTW, island-ness diagnostics). The underlying chunk-level embeddings are not redistributed — they can be regenerated from arXiv source text using the public BGE model and the chunking pipeline at https://github.com/sashakolpakov/dire-rapids-arxiv.

Source-text policy

This release contains only:

• arXiv descriptive metadata (titles, identifiers, categories, chunk counts) — arXiv lists this as CC0-licensed descriptive metadata
• numerical research artifacts (mean-pooled embeddings, layouts) produced by the authors

It does not contain raw arXiv text, PDFs, TeX source, chunk text, or anything otherwise covered by the licenses of the original papers. Users who want the full text of any individual paper should retrieve it directly from arXiv, subject to that paper's license.

Licensing

• The numerical research artifacts in this release (embeddings, layouts) are released under CC-BY 4.0.
• arXiv descriptive metadata fields included here are CC0 per arXiv API terms of use.
• This is not legal advice; users are responsible for compliance with the licenses of any arXiv paper they retrieve separately from this release.

Citation

If you use this dataset, please cite both the paper and the dataset:

@article{kolpakov2026dire,
  author  = {Kolpakov, Alexander and Rivin, Igor},
  title   = {DiRe: Topology-Faithful Dimensionality Reduction},
  journal = {Proceedings of the National Academy of Sciences (submitted)},
  year    = {2026}
}

@dataset{kolpakov_rivin_dire_arxiv_2026,
  author    = {Kolpakov, Alexander and Rivin, Igor},
  title     = {{arXiv paper embeddings for DiRe topology-faithful dimensionality reduction}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19837856}
}

(The Zenodo DOI 10.5281/zenodo.19837856 is the concept DOI: it always resolves to the latest version of this dataset. The version-specific DOI for v1 is 10.5281/zenodo.19837857. Cite the concept DOI in publications. This Hugging Face copy is a convenience mirror.)

Versioning

• v1 (2026-04): Initial release. Pooled embeddings + minimal arXiv metadata + full layout sweep.

Future versions may add: enriched metadata (abstracts, authors, dates, license), chunk-level embeddings, additional reducer baselines.