rwitz/nasa-ntrs-ocr

NASA Technical Reports (NTRS) — modern OCR of the Apollo-era archive

9,000 public NASA technical documents (498,206 pages, 1960–1975) from the NASA Technical Reports Server (NTRS), re-transcribed page by page with lightonocr-3-0.8b, a vision-language OCR model. Tables come out as HTML tables, equations and Greek symbols as Unicode/LaTeX, and headings as Markdown. NASA's own full-text files are not redistributed here (NASA asks that people link to the NTRS record instead), so every row links to its authoritative NTRS citation.

Useful for: pretraining and fine-tuning on technical and scientific text, OCR benchmarking , retrieval and RAG over aerospace history, Apollo/Space Race research, and document-understanding work.

Quick start

from datasets import load_dataset

ds = load_dataset("rwitz/nasa-ntrs-ocr", split="train")
row = ds[0]
print(row["title"], row["year"], row["ntrs_url"])
print(row["text"][:1000])          # new OCR, Markdown
row["cover"]                       # first-page thumbnail (PIL image)

Rights, provenance and what was included

Only documents NASA marks as publicly distributable were used, and only NASA's own report series. Each document was kept only if its NTRS record satisfied all of the following (checked automatically from the record's metadata at crawl time):

• Distribution: PUBLIC (nothing restricted, no CUI, no export control).
• Copyright determination: GOV_PUBLIC_USE_PERMITTED ("Work of the US Gov., public use permitted").
• Not flagged as owned by a contractor, author or publisher, and no third-party material.
• Document type is one of NASA's own series: Technical Memorandum, Technical Note, Technical Report, Technical Publication, Special Publication.
• A PDF is attached.

Deliberately excluded: contractor reports, journal reprints, conference papers, translations of foreign works, anything with an OTHER or any non-government rights label, and anything not PUBLIC.

NASA's terms state that its documents are not protected by copyright unless noted and may be reproduced and distributed without further NASA permission, while contractor/third-party works may be copyrighted. The filter above is designed to keep only the former. Every row carries its rights label and a link (ntrs_url) to the authoritative NTRS record so each item can be verified.

Attribution: Data provided by NASA Scientific and Technical Information Program (https://ntrs.nasa.gov/api/openapi/). NASA is the source of the underlying documents. This dataset is not produced, endorsed or reviewed by NASA, and NASA names/logos are not used to imply endorsement. Opinions in the documents are the authors'.

License of this dataset: the OCR text, metadata and derived fields added here are released under CC0 1.0. The underlying documents are works of the US Government as labelled by NASA. This is a good-faith compliance effort, not legal advice. NTRS rights labels on older records were migrated in bulk, so occasional mislabels are possible.

Availability changes: NASA's Redistributions endpoint is checked at least weekly. Records NASA withdraws or restricts are removed from this dataset.

Takedown / corrections: if you believe any item should not be here, open a discussion on this dataset's Community tab with the id and it will be removed promptly.

Contents

Column	Description
id	NTRS document ID
title, abstract, date, year, sti_type, center, authors, subjects	NTRS metadata
rights	NTRS copyright determination (always GOV_PUBLIC_USE_PERMITTED)
ntrs_url, pdf_url	Source record and original PDF
text	New OCR of the whole document (Markdown/HTML tables), pages joined
pages	JSON list of {page, text, flags} per page
n_pages, n_suspect	Page count; pages flagged as possibly bad (hit token limit or looped)
cover	First-page thumbnail
model	OCR model used

How it was made

1. Crawled the NTRS search API month by month and applied the rights filter above.
2. Downloaded each PDF, rendered every page (longest edge 1540 px) and ran lightonocr-3-0.8b locally through LM Studio.
3. Cleaned output (dropped figure placeholders and model captions, retried loops), kept per-page flags.

Pipeline code: see the repository files ntrs_crawl.py, ntrs_ocr.py, sync.py.

Limitations

• OCR is machine output and unverified. Handwriting, dense fold-out tables, and pages that are mostly figures are the weakest cases. Figures are not transcribed; captions are.
• Page 1 of each document is stored as a small thumbnail only. Full page images are not included; use pdf_url.
• Some pages may have errors, dropped or reordered text; check n_suspect and pages[].flags.

Citation

@dataset{ntrs_lightonocr_ocr,
  title  = {NASA Technical Reports (NTRS) OCR, Apollo era},
  author = {rwitz},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rwitz/nasa-ntrs-ocr},
  note   = {Source documents: NASA Technical Reports Server, https://ntrs.nasa.gov/}
}