vladimirbesk/tsiolkovsky-papers

Tsiolkovsky Papers: the complete personal archive as a machine-readable corpus

Machine transcriptions of all 51,008 sheets of fond 555 of the Archive of the Russian Academy of Sciences — the personal archive of Konstantin Tsiolkovsky (1857–1935), who derived the rocket equation and described the multistage rocket decades before anyone could test either.

The archive had been scanned and put online, but without a catalogue you could query, full-text search, or a dataset. This is that dataset.

Canonical archive of record: Zenodo, DOI 10.5281/zenodo.21705221. This repository is a convenience copy for machine-learning work.

What makes it unusual

Transcription accuracy of this corpus is measured, not asserted — which is rare for handwritten material, because measuring it normally requires ground truth that archives do not have.

Sheets	51,008
Archival files	2,019 (complete fond)
Handwritten / typed / notes	34,903 / 14,585 / 1,520
Uncertainty marks	299,939
Authorial deletions	36,446
Dated files	1,969 (1878–1935)

Fields

Field	Meaning
id	opis/delo/sheet, unique
opis, delo, sheet	archival inventory, file and sheet number
text	the transcription, with markup preserved
page_type	hand, typed or note (80% accurate, see limits)
words, chars	length
uncertain_marks	count of places the model marked as doubtful
struck_out	passages struck out by the author
title, year_from, year_to, date_conjectural	from the archival description
scan_url	the page for this file on the archive's own site

Markup inside text: слово[?] an uncertain reading, [неразборчиво] illegible, ~~слово~~ struck out by the author, [вставка: …] an insertion, [на полях: …] a marginal note, [другой почерк: …] a different hand. Original orthography is preserved, including pre-reform letters ѣ, і, ъ, ѳ.

Measured accuracy

Against printed editions, where the archival file holds the same redaction that was published: 98.1% of characters on typescript, 81.1% on handwriting (scored on the part read by the earlier pipeline); 92.3% across ten files of the finished corpus. Where a file holds a draft rather than the printed text, agreement drops to about 24% — that is the distance between manuscript and edition, not misreading, and the two are reported separately rather than averaged.

Without ground truth. Part of the fond holds one text twice, as the author's autograph and as a typed copy of it, in the same file. Reading both isolates the difficulty of the hand, since source and pipeline are identical. Over 1,759 such pairs from 224 files, two readings of one text agree on a median 37% of words, with a longest verbatim run of 10 words.

That estimate was checked against real ground truth on files that also have a printed edition: unbiased to within a percentage point, with rank correlation 0.67 over all 55 pairs and 0.92 over the 32 where the edition is demonstrably the same text.

Method: arXiv:2608.03617.

Limits — please read before using

No human has checked this against the scans. Not one sheet. It is a machine transcription with its doubts marked, not a scholarly edition. Good for search, retrieval, dating and training; not for quotation without checking the scan, and every record carries scan_url for that.

Redactions cannot be collated at this quality. Two variants of one work in this fond share 19% of their words — below the rate at which two readings of one page agree. Authorial revision cannot be told from misreading here.

page_type is 80% accurate, not perfect: it was validated on 19 hand-labelled pages, passed all 19, and scored 80% against an independent signal over 4,675 sheets. Carbon copies and faded typescript read as handwriting to it.

Uncertainty marks are a feature. 299,939 of them, about six per sheet. A transcription whose doubtful places are visible can be checked; a fluent one whose doubtful words were invented cannot.

Two sheets are a German typewritten review that every model refused to reproduce verbatim; they were read another way and say so in their own text.

Suggested uses

Training and evaluating handwritten text recognition for Russian, including pre-reform orthography. Benchmarking uncertainty calibration: the marks give a model's own doubt, and the autograph/typescript pairs give an external check on it. Text reuse detection across a single author's archive. Historical research on where the fond has been unreachable for lack of search.

Citation

@dataset{beskorovainyi_tsiolkovsky_2026,
  author    = {Beskorovainyi, Vladimir},
  title     = {Tsiolkovsky Papers: a machine-readable catalogue of fond 555,
               Archive of the Russian Academy of Sciences},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21705221}
}

Released CC0: use without permission and without attribution, though attribution is welcome. The scans themselves remain with the archive and are not redistributed here; scan_url points to them.

Independent project, not affiliated with or endorsed by the Archive of the Russian Academy of Sciences.