LICENSING OF THIS DATASET ========================= This release combines papers with different licences and a CC0 organisational layer. The compilation licence does not relicense the papers. The nine configs are sample, metadata, versions, files, paper_text, latex, source, pdf and ps. Provenance is recorded in columns, chiefly in files; there is no separate provenance config. 1. PAPERS AND CONTRIBUTED CONTENT -------------------------------- Rights in the papers and submitted files remain with their respective holders, subject to the applicable licence or public-domain dedication. This dataset grants no additional rights to that content. This includes full text in paper_text and sample as well as files in latex, source, pdf and ps. metadata.license records the paper-level licence URL supplied by arXiv's harvested metadata, or NULL if none is recorded. The value is copied into versions, paper_text and sample and agrees on every shared paper. The four file-content configs have no licence column; join to metadata on paper_id. The value in versions is copied from the paper record, not independently recorded for each revision. arXiv allows different versions to have different licences. Check the relevant version when relying on its terms. The dataset column is an index to the recorded licence, not a replacement for those terms. See arXiv's licensing policy: https://info.arxiv.org/help/license/index.html Counts below are per paper, out of 3,148,796: papers share recorded licence --------- ----------- ----------------------------------------------- 1,902,375 60.4159% arXiv perpetual non-exclusive licence 1.0 572,219 18.1726% CC BY 4.0 452,715 14.3774% none recorded (NULL) 88,143 2.7993% CC BY-NC-ND 4.0 65,301 2.0738% CC BY-NC-SA 4.0 30,187 0.9587% CC BY-SA 4.0 21,599 0.6859% CC0 1.0 7,912 0.2513% CC BY 3.0 5,871 0.1865% CC BY-NC-SA 3.0 2,474 0.0786% CC public domain --------- ----------- 3,148,796 100.0000% nine declarations, plus missing values Creative Commons licences and public-domain declarations account for 793,706 papers (25.21%); the other 2,355,090 (74.79%) have the arXiv non-exclusive licence or no recorded licence. NonCommercial terms occur in 159,315 papers, NoDerivatives in 88,143 and ShareAlike in 101,359; these groups overlap. The arXiv non-exclusive licence grants arXiv permission to distribute a paper. It does not itself grant downstream redistribution rights: https://arxiv.org/licenses/nonexclusive-distrib/1.0/ A NULL value does not establish that a paper is in the public domain. Most NULL values occur in older records. In SQL, use IS NULL to select them; inequality comparisons exclude NULLs. For a selection of CC BY, CC BY-SA and public-domain declarations: SELECT paper_id FROM 'hf://datasets/secemp9/arxiv-complete/metadata/*.parquet' WHERE license LIKE 'http://creativecommons.org/licenses/by/%' OR license LIKE 'http://creativecommons.org/licenses/by-sa/%' OR license LIKE 'http://creativecommons.org/publicdomain/%' OR license = 'http://creativecommons.org/licenses/publicdomain/' This selects 634,391 papers (20.15%). Applicable conditions, including attribution and ShareAlike where present, still apply. Licence shares differ by population. NULLs are 452,715 of 3,148,796 metadata papers (14.38%), 424,782 of 2,856,227 paper_text rows (14.87%), and 649,392 of 5,030,612 version rows (12.91%). The values agree; the denominators differ. 2. COMPILATION AND METADATA -------------------------- The organisational layer -- the file inventory, SHA-256 digests, sizes and paths, version history and flags, the provenance columns, and the schema tying them together -- is released under CC0 1.0, to the extent it is protectable at all. Facts about which files exist and what they hash to are not creative works, and no attempt is made to claim them. CC0 1.0: https://creativecommons.org/publicdomain/zero/1.0/ Bibliographic metadata originates from arXiv's OAI-PMH interface. arXiv's licensing policy separately places its metadata under CC0, including the metadata fields copied into this release. That metadata policy and this compilation dedication do not extend to the papers' full text or submitted files. 3. CONTENT HANDLING ------------------- No missing PDF was generated from TeX or another representation. paper_text is derived by assembling and decoding source files, as documented in the README. Individual file bytes were preserved apart from these operations: Three corrupt mirrored PDFs were replaced with complete arXiv copies: 2401.18030 v3 0911.0496 v1 1001.4505 v1 The first mirrored object was a 267-byte HTML redirect to browse.arxiv.org/pdf/2401.18030v3. Its recorded bytes reproduce the repair journal's SHA-256 prefix 2e066757... and the mirror manifest's MD5 prefix db732060.... No published column marks these repairs: source remains gcs, while size, sha256 and versions.pdf_sha256 describe the replacement bytes. The internal repaired_from field was not published. The three repairs came from 54 flagged candidates. The comparison retained 42 defective objects whose bytes matched arxiv.org, replaced three, identified eight intact PNG/JPEG figures named .pdf in ps, and excluded one partial temporary file that was never published. Thus 45 candidates were defective PDFs, not 54. Two of the 42 retained objects have web_scrape provenance. The README's "The 54 flagged PDF candidates" section gives the evidence and explains the comparison script's invalid URL requests for the eight figures. Gzip transport wrappers were decompressed for 2,168,515 ps-kind objects: 2,070,188 named .ps.gz, 90,409 .eps.gz and 7,918 with other names. Original names, sizes and digests remain in files.stored_as, stored_size and stored_sha256. Those columns are not present in the ps content config. 4. ATTRIBUTION AND TAKEDOWN -------------------------- Cite the papers your work uses. metadata.arxiv_abs_url and versions.arxiv_pdf_url provide arXiv record and download links; neither column is null. Authors who want their work removed can open a discussion on the dataset repository. Their work will be removed from the next revision. This does not perform an arXiv withdrawal; that process must be pursued with arXiv directly.