Scientific-Summaries

22 million LLM-generated structured summaries of scientific papers, enriched with OpenAlex scholarly metadata. Each paper has an 18-field structured summary covering methodology, key results, claims, limitations, and more. This public dataset includes full paper text for ~5.3 million papers where open-access status has been confirmed -- either through OpenAlex metadata or because the paper originates from a permissively licensed source such as the arXiv preprint server, bioRxiv, medRxiv, or ChemRxiv.

Note: The 22M count refers to paper summaries -- every paper has a complete structured summary regardless of whether the full text is included. Full text is shared only for papers with confirmed open-access status. A private version with full text for all 22M papers is available for research collaborators.

Overview

Full Text Policy

Full text (text_sanitized, text_raw) is included in this public dataset when either of the following is true:

1. The paper originates from a permissively licensed source: arXiv preprint server, bioRxiv, medRxiv, ChemRxiv, PsyArXiv, EdRxiv, MetaArXiv, or NeurIPS proceedings. These are all open-access by design.
2. The paper is flagged as open access by OpenAlex: For papers from PubMed and the general web-papers collection, we include full text only when OpenAlex's oa_is_oa field confirms open-access status.

Papers without full text still have complete 18-field structured summaries and all available OpenAlex metadata (DOI, citation count, author affiliations, referenced works, etc.).

---

Bias, Risks, and Limitations

Data Biases

• The corpus reflects the linguistic and disciplinary biases of its sources: predominantly English-language, Western-institution-affiliated research. Geographic and linguistic underrepresentation in the underlying scientific literature is inherited by the dataset.
• LLM-generated summaries may inherit biases from both the summarization model (Gemini 2.0 Flash Lite) and the source literature, including framing biases, omission of minority perspectives, and overrepresentation of dominant research paradigms.
• Disciplinary coverage is uneven: arXiv (computer science, physics, mathematics) and PubMed (biomedical sciences) are well-represented, while humanities, social sciences, and non-English-language scholarship are underrepresented.
• The web-papers subset drawn from S2ORC may overrepresent digitally accessible and English-language publications, further amplifying existing inequities in scientific publishing.

Data Limitations

• Summary quality is bounded by the LLM used for generation (Gemini 2.0 Flash Lite): factual errors, hallucinations, and omissions may propagate into structured fields without explicit indication. Summaries should be treated as derived representations, not ground-truth substitutes for original papers.
• Full paper text is available for only ~5.3 million of the 22 million papers (~24%); the remaining papers have summaries and metadata only.
• Summaries were generated in a single pass without iterative verification or large-scale human spot-checking, limiting the ability to quantify systematic failure modes across disciplines.
• The corpus may contain near-duplicate content with evaluation benchmarks used in the wider literature; downstream users training LLMs should conduct independent decontamination.
• OpenAlex metadata matching quality varies; not all papers have complete enrichment, and match accuracy depends on DOI availability and title-matching heuristics.
• Pre-training experiments reported in the paper are limited to a 1.7B parameter model at 15B tokens, representing an early-stage probe rather than a definitive assessment at frontier model scales.

Recommendations

Users should treat structured summaries as derived, approximate representations rather than authoritative records of the original papers, particularly in high-stakes domains such as clinical medicine or policy analysis. The is_retracted flag from OpenAlex should be used to filter discredited findings. The ethical_considerations field in each summary surfaces author-reported and model-identified concerns at the paper level.

---

Uses

Direct Use

• Retrieval-augmented generation (RAG) over scientific literature at scale, with fine-grained field-level retrieval across methodology, results, claims, and limitations via 42 pre-built search indices.
• LLM pretraining with high-density structured knowledge injection; controlled ablations show structured summaries outperform raw science papers and Wikipedia at matched token budgets on knowledge-intensive benchmarks (e.g., MMLU).
• Scientific claim verification and evidence synthesis using the structured claims field (JSON array of claims with supporting and contradicting evidence).
• Automated literature review and systematic review assistance, leveraging the 18-field taxonomy to identify methodology, key results, and limitations across papers.
• Cross-disciplinary knowledge synthesis and hypothesis generation across 22M papers spanning all major scientific disciplines.
• Bibliometric analysis using OpenAlex enrichment metadata (citation counts, author affiliations, referenced works, language, retraction status).

Out-of-Scope Use

• Direct clinical or policy decision-making based solely on LLM-generated summaries without consulting original papers -- summaries may contain hallucinations or omissions.
• Reproducing or substituting original papers -- structured summaries capture distilled factual content, not the full nuance of the original work.
• Training models for high-stakes applications without independent verification of summary accuracy and decontamination against evaluation benchmarks.

---

Personal and Sensitive Information

• The dataset contains publicly available scholarly metadata: author names, institutional affiliations, DOIs, and citation information sourced from OpenAlex. This information is already part of the public scholarly record.
• No private personal data is collected beyond what is available in the public scholarly record. Author names and affiliations are retained to ensure proper credit attribution, consistent with standard academic citation practice.
• The ethical_considerations field in each summary may surface author-reported ethical concerns from the original papers, which could reference sensitive topics (e.g., human subjects research, clinical trials). These reflect the content of the source papers, not independently collected personal data.

---

Social Impact

This dataset democratizes access to structured representations of 22 million scientific papers across all major disciplines, lowering barriers to scientific retrieval and enabling researchers to search across methodology, claims, and limitations rather than just titles and abstracts.

Positive impacts: - Supports AI-assisted research and discovery workflows, potentially accelerating scientific progress. - Enables efficient navigation of the scientific literature for researchers in under-resourced institutions. - Provides structured signals (claims, limitations, ethical considerations) that improve the granularity and faithfulness of retrieval-augmented generation systems.

Risks and concerns: - May amplify existing inequities in scientific publishing by reflecting the overrepresentation of English-language and Western-institution research in the underlying corpora. - Structured summaries should not replace reading original papers, particularly in high-stakes domains such as clinical medicine, pharmacology, or public policy. - The ethical_considerations field and is_retracted flag provide guardrails for responsible downstream use, but users must exercise independent judgment about data quality and applicability.

---

Synthetic Data Statement

All 22,087,912 structured summaries in this dataset are LLM-generated (synthetic), produced using Gemini 2.0 Flash Lite in a single-pass extraction from paper full texts. The dataset represents approximately 60 billion tokens of synthetic structured data when tokenized.

• The 18-field structured summaries are entirely synthetic: they are machine-generated representations of paper content, not human-authored descriptions.
• The underlying paper texts (available for 5.3M open-access papers) are not synthetic -- they are the original human-authored scientific papers.
• OpenAlex metadata (author names, citations, DOIs, etc.) is not synthetic -- it is sourced from the OpenAlex scholarly database.
• Summarization metadata (summary_model, summary_tokens_in, summary_tokens_out) is retained for every paper to support reproducibility and quality auditing.

---

Search Infrastructure

This dataset includes a comprehensive search infrastructure with 42 pre-built indices for both semantic (FAISS) and keyword (BM25) search across all papers, text chunks, and claims.

See SEARCH_USAGE.md for complete documentation, code examples, and deployment guides.

Search Capabilities at a Glance

Quick Start: Semantic Search

pip install faiss-cpu numpy pandas pyarrow huggingface_hub sentence-transformers

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

# Download
snapshot_download("laion/Scientific-Summaries", repo_type="dataset",
    local_dir="scientific-summaries",
    allow_patterns=[
        "search/faiss/summaries/concatenated_summary.faiss",
        "search/faiss/summaries/concatenated_summary.faiss.ivfdata",
        "search/mappings/paper_id_order.parquet",
    ])

# Search
model = SentenceTransformer("Alibaba-NLP/gte-modernbert-base")
query_vec = model.encode(["CRISPR gene editing efficiency"],
                         normalize_embeddings=True).astype(np.float32)
index = faiss.read_index("scientific-summaries/search/faiss/summaries/concatenated_summary.faiss")
index.nprobe = 32
scores, indices = index.search(query_vec, 10)

# Results
paper_order = pd.read_parquet("scientific-summaries/search/mappings/paper_id_order.parquet")
for idx, score in zip(indices[0], scores[0]):
    if idx >= 0:
        row = paper_order.iloc[idx]
        print(f"[{score:.4f}] {row['subset']}: {row['paper_id'][:80]}")

Quick Start: Keyword Search

pip install bm25s PyStemmer pandas pyarrow huggingface_hub

import bm25s, Stemmer, pandas as pd
from huggingface_hub import snapshot_download

# Download
snapshot_download("laion/Scientific-Summaries", repo_type="dataset",
    local_dir="scientific-summaries",
    allow_patterns=[
        "search/bm25/summaries/executive_summary/*",
        "search/mappings/paper_id_order.parquet",
    ])

# Search
stemmer = Stemmer.Stemmer("english")
query_tokens = bm25s.tokenize(["CRISPR gene editing efficiency"], stemmer=stemmer)
retriever = bm25s.BM25.load("scientific-summaries/search/bm25/summaries/executive_summary")
results, scores = retriever.retrieve(query_tokens, k=10)

# Results
paper_order = pd.read_parquet("scientific-summaries/search/mappings/paper_id_order.parquet")
for doc_idx, score in zip(results[0], scores[0]):
    row = paper_order.iloc[doc_idx]
    print(f"[{score:.2f}] {row['subset']}: {row['paper_id'][:80]}")

Compact zvec BM25 Indices

Self-contained BM25 search indices built with zvec — no external BM25 library needed. Separate text sidecars enable fast search with optional text retrieval. Total: ~144 GB (vs 313 GB for bm25s indices).

Architecture: Search returns IDs → retrieve text from sidecars. Claims can reference both claims_text (for claim text) and summaries_text (for full paper context).

See search/zvec/USAGE.html for complete usage instructions.

Quick-start (summaries search):

pip install zvec PyStemmer pyarrow huggingface_hub

import hashlib, json, re, struct
from collections import Counter
import Stemmer, zvec
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download
from pathlib import Path

# Download index + vectorizer + text sidecar (~107 GB total, or 35 GB without text)
snapshot_download("laion/Scientific-Summaries", repo_type="dataset",
    local_dir="scientific-summaries",
    allow_patterns=["search/zvec/summaries_bm25/**",
                    "search/zvec/summaries_text/**",
                    "search/zvec/summaries_bm25_vectorizer.json"])

# Tokenizer (must match index build)
_stemmer = Stemmer.Stemmer("english")
_tok_re = re.compile(r"[a-zA-Z0-9]+")
HASH_SPACE = 2**20

def tokenize(text):
    return _stemmer.stemWords(_tok_re.findall(text.lower()))
def token_to_id(t):
    return struct.unpack("<Q", hashlib.md5(t.encode()).digest()[:8])[0] % HASH_SPACE

# Load vectorizer and collection
with open("scientific-summaries/search/zvec/summaries_bm25_vectorizer.json") as f:
    d = json.load(f)
idf = {int(k): v for k, v in d["idf"].items()}
col = zvec.open("scientific-summaries/search/zvec/summaries_bm25")

# Build text sidecar index (one-time, fast)
sidecar_path = Path("scientific-summaries/search/zvec/summaries_text")
text_index = {}
for pf in sidecar_path.glob("*.parquet"):
    table = pq.read_table(pf, columns=["paper_id"])
    for i, pid in enumerate(table.column("paper_id")):
        text_index[str(pid.as_py())] = (str(pf), i)

# Search
query = "CRISPR gene editing efficiency"
tf = Counter(token_to_id(t) for t in tokenize(query))
qvec = {tid: idf[tid] * c for tid, c in tf.items() if tid in idf and idf[tid] > 0}

results = col.query(
    vectors=zvec.VectorQuery(field_name="bm25", vector=qvec), topk=5,
    output_fields=["paper_id"])

# Retrieve text from sidecar
for doc in results:
    paper_id = doc.field("paper_id")
    if paper_id in text_index:
        pf_path, row_idx = text_index[paper_id]
        table = pq.read_table(pf_path)
        summary = str(table.column("concatenated_summary")[row_idx].as_py())
        print(f"[{doc.score:.4f}] {summary[:200]}...")

For claims search: Download search/zvec/claims_bm25/, search/zvec/claims_text/, and search/zvec/claims_bm25_vectorizer.json. Claims can reference both claims_text (for claim text) and summaries_text (for full paper summary).

Performance Benchmarks

Retrieval latency for top-k=10 queries (5 test queries, single-threaded, cold start):

Key insights: - Claims search is faster (1.8s vs 2.7s) due to smaller per-vector overhead - Claim text retrieval is very fast (92ms) from the compact 4 GB sidecar - Summary text retrieval (3.7-3.8s) is the bottleneck — 72 GB parquet files - Full claims pipeline (search + claim text + paper summary) completes in ~5.7s - Text sidecar architecture enables search-only queries (<2s) without loading text

Benchmark script: benchmark_bm25_retrieval.py

Search Index Statistics

Source text for displaying results: summary parquets (~346 GB), chunk parquets (~64 GB), claim parquets (~5 GB).

Example Scripts

Technical Details

• Embedding model: Alibaba-NLP/gte-modernbert-base (768 dimensions)
• FAISS index type: IVF with PQ48 compression (768 dims to 48 bytes, 16x compression)
• BM25 library: bm25s (Lucene variant, English stemmer)
• Sharding: Large BM25 indices split into 5M-document shards for memory efficiency

---

Subsets in Detail

arXiv (3.3M summaries, ~100% with full text)

Near-complete coverage of arXiv as of April 2026 (99.8% of all 3,011,619 papers). Includes papers from all arXiv categories: cs, math, physics, q-bio, q-fin, stat, eess, econ. All arXiv papers are permissively licensed, so full text is included for all.

Papers come from three sources that were merged and deduplicated: - Original S2ORC/GROBID-parsed arXiv papers (~1M) - Newly downloaded and PyMuPDF-parsed arXiv PDFs from Google Cloud Storage (~1.6M) - arXiv papers identified within the general S2ORC papers-1 collection via title matching (~735K)

Bethgelab (335K summaries, 100% with full text)

Parsed papers from the bethgelab/paper_parsed_jsons dataset. All papers come from permissively licensed preprint servers, so full text is included for all.

PubMed (1.7M summaries, ~16% with full text)

Biomedical papers from PubMed Central, parsed from the S2ORC collection. Full text is included only for papers confirmed as open access by OpenAlex (~270K papers). All 1.7M papers have complete summaries and metadata regardless.

Web Papers (16.7M summaries, ~8% with full text)

General scientific papers from the S2ORC (Semantic Scholar Open Research Corpus) papers-1 collection, covering all scientific disciplines. This is the largest subset. Full text is included only for open-access papers confirmed by OpenAlex (~1.4M papers). All 16.7M papers have complete summaries and metadata.

The web-papers subset was created in collaboration with Wynd Labs, who supported the large-scale web scraping and paper collection effort that made this dataset possible.

Schema (44 columns)

Summary Fields (18 LLM-generated fields)

Each paper has a structured summary generated by priv-gemini-2.0-flash-lite:

Source Metadata

paper_id, subset, source_title, source_authors, source_year, source_venue, source_doi

OpenAlex Enrichment

openalex_id, oa_doi, oa_title, oa_type, oa_year, oa_cited_by_count, oa_language, oa_is_retracted, oa_is_open_access, oa_authorships (JSON), oa_referenced_works (JSON), match_method, has_openalex_match

Text Fields

Summarization Metadata

summary_model, summary_tokens_in, summary_tokens_out

Usage

from datasets import load_dataset

# Load arXiv summaries (3.3M papers, all with full text)
ds = load_dataset("laion/Scientific-Summaries", "arxiv")

# Load web papers (16.7M summaries, ~8% with full text)
ds = load_dataset("laion/Scientific-Summaries", "web-papers")

# Access a summary
paper = ds['train'][0]
print(paper['executive_summary'])

# Access structured claims
import json
claims = json.loads(paper['claims'])
for claim in claims:
    print(f"Claim: {claim['description']}")

# Filter for papers with full text
with_text = ds['train'].filter(lambda x: x['text_sanitized'] is not None)

# Filter by field
cs_papers = ds['train'].filter(
    lambda x: x['field_subfield'] and 'Computer Science' in x['field_subfield']
)

Project Alexandria

This dataset is part of Project Alexandria, an initiative to create comprehensive, machine-readable representations of the world's scientific literature. The project aims to summarize every scientific paper ever published and enrich it with structured metadata.

Current progress: 22M+ papers summarized, covering ~99.8% of arXiv, major portions of PubMed, and millions of general scientific papers across all disciplines.

Acknowledgments

A huge thank you to our collaborators and data sources:

• Wynd Labs for their invaluable collaboration on large-scale web paper collection and scraping infrastructure that made the web-papers subset possible
• OpenAlex for scholarly metadata enrichment (DOIs, citation counts, author affiliations, open-access status)
• Semantic Scholar / S2ORC for GROBID-parsed paper collections
• bethgelab for curated preprint server paper collections
• arXiv for open access to preprints via Google Cloud Storage

Citation

@misc{alexandria2025,
    title={Project Alexandria: Structured Summaries of Scientific Papers at Scale},
    author={REDACTED},
    year={2025},
    eprint={2502.19413},
    archivePrefix={arXiv}
}

License

CC-BY-4.0