Datasets:
Title stringlengths 6 239 | Raw-Text stringlengths 5.47k 1.96M | Clean-Text stringclasses 0
values | Hash stringlengths 64 64 |
|---|---|---|---|
CAST: Causal Advantage-Structured Training with Spatially Grounded Compositional Rewards for Diffusion Models | CAST: Causal Advantage-Structured Training with Spatially Grounded Compositional Rewards for Diffusion Models
Shu Yu
Affiliation: Shanghai Artificial Intelligence Laboratory, Shanghai, China
Affiliation: Shanghai Innovation Institute, Shanghai, China
Affiliation: Fudan University, Shanghai, China
Email: yushu@pjl... | null | 4915b0055d1518b8951472f13c5850aa5fa451314f0a042ed052880c1b2ecb4e |
Conformalized Super Learner | Conformalized Super Learner
Zhanli Wu
Affiliation: Department of Mathematics
Affiliation: King’s College London
Affiliation: London, UK
Email: zhanli.wu@kcl.ac.uk
Fabrizio Leisen
Affiliation: Department of Mathematics
Affiliation: King’s College London
Affiliation: London, UK
Email: fabrizio.leisen@kcl.ac.u... | null | 513723d8b4d21fd9243fa813080f758d110f156f65e925946ba1ce7967391076 |
Tensor Network Moral Graph Recovery of Discrete Probability Distributions | "Álvaro Troyano et al\n\nTensor Network Moral Graph Recovery of Discrete Probability Distributions\(...TRUNCATED) | null | 2e72ca4f0537c4720de6846a5008d4747660b7f6f2706c66e6c231de7bde6c47 |
AST-Based Automated Elimination of break and continue Statements in Java Code | "AST-Based Automated Elimination of break and continue Statements in Java Code\n\n\nAndrés Juárez\(...TRUNCATED) | null | 97261e907b6ce6e17e685514c3d8f82880c4c3a75bf9b3a30f001650fe8a527e |
Optimal Network Dependence in Distributed Stochastic Optimization via Tree Routing | "Optimal Network Dependence in Distributed Stochastic Optimization via Tree Routing\n\n\nRunze You\n(...TRUNCATED) | null | c6e929ca0c251555dff7272a91cabcfd5818cabadb21abd7534467993f54f196 |
Single-shot coherent process tomography and mid-infrared polarimetry with undetected photons | "Single-shot coherent process tomography and mid-infrared polarimetry\nwith undetected photons\n\n\n(...TRUNCATED) | null | 8544413d1eb4e180054c06df7257df663b95ae943f4794cf31d50e4865893315 |
"InfiMed2: A Generalist Medical Multimodal Foundation Model from Contextual Evidence and Stability-A(...TRUNCATED) | "InfiMed2: A Generalist Medical Multimodal\nFoundation Model from Contextual Evidence\nand Stability(...TRUNCATED) | null | a95450e92f0fc737fbdd9de7394f54aa178d40a4c36c8179586f0865bfc64c87 |
Limit Theorems for the Pitman-Yor Frequency Spectrum | "Limit Theorems for the Pitman-Yor Frequency Spectrum\n\n\nRoss Maller\n\n\nSoudabeh Shemehsavar\n(...TRUNCATED) | null | 63cbb1fc828fd03766af1b7b5181cf27ddea3443cf9982170bd17740c5dd068e |
Limit laws for component-pruned sparse random graphs and percolated tori | "Limit laws for component-pruned sparse random graphs and percolated toriThanks: The second author r(...TRUNCATED) | null | 05f1755e186a14e76a5e306c259c5ec9b81a8b871ea4eedfed6a746ff624214b |
Platform for Representation and Integration of multimodal Molecular Embeddings | "Platform for Representation and Integration of multimodal Molecular Embeddings\n\n\nErika Yilin Zhe(...TRUNCATED) | null | a981a562f08b6559a766dd6c492e21c2da060d84e4ff06bfda3632542d37ef55 |
ArXiv 20K Title-to-Paper
A mixed-domain collection of full scientific paper text extracted from arXiv.
The dataset is designed primarily for experiments involving title-conditioned scientific text generation, long-form language modeling, document modeling, and scientific-text research.
Dataset Summary
The initial collection contained 20,000 papers sampled across a broad mixture of arXiv subject categories.
A conservative structural filtering pass removed obviously broken documents, severe extraction failures, extreme repetition, error pages, and other unusable records.
Final row count in this release: 19,822 papers.
The filtering pass retained 99.11% of the original collection.
This release preserves both the extracted source text and a dedicated column for future cleaned text.
Schema
The dataset contains four columns:
| Column | Description |
|---|---|
Title |
Title of the arXiv paper |
Raw-Text |
Full text extracted from the arXiv HTML rendering with only minimal extraction normalization |
Clean-Text |
Reserved for the cleaned version of the paper text |
Hash |
SHA-256 hash of the raw extracted text |
Current Clean-Text status
In the initial release, Clean-Text is null.
The raw text is intentionally preserved before a more aggressive corpus-wide cleaning pass is performed.
A later revision is intended to populate this column while retaining Raw-Text unchanged, allowing direct comparison between the source extraction and cleaned representation.
Example
from datasets import load_dataset
dataset = load_dataset(
"Ma7ee7/ArXiv-20K-Title-to-Paper",
split="train",
)
row = dataset[0]
print(row["Title"])
print(row["Raw-Text"][:2000])
print(row["Clean-Text"])
print(row["Hash"])
A title-to-paper training pair can simply be constructed as:
input_text = row["Title"]
target_text = (
row["Clean-Text"]
if row["Clean-Text"] is not None
else row["Raw-Text"]
)
Collection
Papers were sampled from a mixture of arXiv categories rather than a single research field.
The category pool included areas from:
- computer science
- machine learning
- statistics
- mathematics
- physics
- electrical engineering
- quantitative biology
- economics
- quantitative finance
Example categories include:
Computer Science
cs.AIcs.CLcs.CVcs.LGcs.NEcs.ROcs.IRcs.SEcs.CRcs.HCcs.DS
Statistics
stat.MLstat.THstat.APstat.CO
Mathematics
math.PRmath.STmath.OCmath.NAmath.COmath.LOmath.IT
Physics
physics.comp-phphysics.data-anquant-phcond-mat.stat-mechhep-thastro-ph.CO
Electrical Engineering
eess.SPeess.SYeess.IV
Quantitative Biology
q-bio.NCq-bio.QMq-bio.BM
Economics and Quantitative Finance
econ.EMecon.THq-fin.STq-fin.EC
A per-category cap was used during collection to reduce the chance that a single large field would dominate the corpus.
Text Extraction
Full paper text was collected primarily from arXiv's HTML representations.
The initial extraction preserves the paper text relatively conservatively.
During extraction:
- website interface elements were removed
- script and style elements were removed
- mathematical expressions were represented using their available LaTeX-style alternative text
- obvious HTML serialization artifacts were normalized
- excessive blank-line runs were collapsed
The resulting text is stored in Raw-Text.
No language model was used to rewrite or paraphrase the scientific content.
Filtering
The original collection contained:
20,000 rows
After filtering:
19,822 rows were retained
178 rows were rejected
This corresponds to a survival rate of approximately:
99.11%
The filtering stage was deliberately conservative.
Documents could be rejected for problems such as:
- extreme repeated-line extraction
- very short text
- implausibly large extraction
- too few words
- severe repeated-line dominance
- apparent error or conversion pages
- malformed records
- exact duplicate text
- duplicate titles
The most common detected failure mode in the initial corpus was extreme line repetition.
Corpus Statistics
For the surviving dataset:
- Rows: 19,822
- Median characters per paper: ~73,603
- Mean characters per paper: ~87,316
- Median words per paper: ~11,718
- Mean words per paper: ~14,239
The shortest retained paper contained approximately 5.5k characters.
The largest retained extraction contained approximately 1.96M characters.
Hashes
Hash is a SHA-256 digest derived from Raw-Text.
The hash is intended to provide:
- exact-text deduplication
- integrity checking
- stable identity for a specific raw extraction
- comparison across future cleaning revisions
The hash is based on the raw text rather than Clean-Text, so future cleaning changes do not alter the identifier.
Intended Uses
Potential uses include:
- title-to-paper generation experiments
- scientific language modeling
- long-context modeling
- scientific document completion
- research on document structure
- scientific-text compression
- retrieval experiments
- pretraining or continued pretraining
- studying relationships between titles and full documents
Limitations
Extraction artifacts
Raw-Text is derived from automatic HTML extraction.
Some papers may still contain artifacts involving:
- tables
- figures
- captions
- mathematical formatting
- references
- unusual Unicode
- repeated fragments
- document-specific layout
The filtering process targets severe failures rather than guaranteeing perfectly clean prose.
Clean-Text
Clean-Text is currently unpopulated.
A later dataset revision is expected to add a deterministic cleaned representation while retaining the original raw extraction.
Sampling
This is not a complete or statistically representative sample of all arXiv papers.
The corpus was intentionally sampled across a selected group of scientific categories to create a diverse experimental dataset.
Scientific accuracy
Inclusion in this dataset does not imply that a paper's claims are correct, peer reviewed, replicated, or endorsed.
Licensing
This dataset aggregates text from papers hosted by arXiv.
Individual papers on arXiv may be distributed under different licenses and copyright terms.
The license: other designation for this dataset does not replace or override the license associated with any individual paper.
Users are responsible for determining whether their intended use complies with the license and rights associated with the underlying works.
Dataset Version
v0.1
Initial filtered release.
- 20,000 initially collected papers
- 19,822 retained after structural filtering
Raw-TextpopulatedClean-Textreserved but currently null- SHA-256 raw-text hashes included
Future revisions may populate Clean-Text and improve quality filtering while preserving the raw source column.
- Downloads last month
- 51