Reubencf/konkani-newspapers

Konkani Newspapers (Vauraddeancho Ixtt, 2012-2016)

Scanned pages of Vauraddeancho Ixtt (Vauraddeancho Ixtt / वावराड्यांचो इष्ट, "Worker's Friend"), a Konkani-language weekly newspaper published from Pilar, Goa, India, and billed on its own masthead as "Goa's only Konkani Weekly since 1933".

This is a page-image dataset, not a text corpus. Every row is one full scanned newspaper page, together with the date and page number parsed from the source filenames. There is no transcription or OCR text: the intended use is as raw material for OCR, layout analysis, and document-understanding work on Konkani, a low-resource Indian language for which very little scanned print data is publicly available.

The paper is set mainly in Romi Konkani (Konkani written in Roman script) with some Devanagari, English, and Portuguese-derived vocabulary. Romi Konkani is substantially under-served by existing OCR systems, which is much of the point of releasing this.

Schema

Column	Type	Description
image	datasets.Image()	The scanned page, as a PNG or JPEG image
date	string	Issue date, ISO YYYY-MM-DD, parsed from the source folder name
year	int32	Year of the issue
month	int32	Month of the issue, 1-12
day	int32	Day of the issue
page	int32	The newspaper's own printed page number, parsed from the source filename. Sparse — see Completeness.
issue	string	Original source folder name, so rows trace back to the source
num_pages	int32	Number of scanned pages present in this dataset for that issue. NOT the number of pages the printed issue had, and usually less than max(page).

Rows are sorted by (date, page) ascending, so the dataset reads in chronological and then page order.

Read the num_pages definition carefully. It counts the images actually available here. For 148 of the 180 issues the scan is partial, so num_pages is lower — often by 3 or 4 — than the highest page number present. It is a measure of what you have, not of how long the newspaper was.

Loading

from datasets import load_dataset

ds = load_dataset("Reubencf/konkani-newspapers", split="train")
print(ds[0]["date"], ds[0]["page"])
ds[0]["image"].show()

# all pages of one issue, in printed page order
issue = ds.filter(lambda r: r["date"] == "2013-04-13")

Note that num_pages is the count of scans available, not the printed length of the issue, and page is often non-contiguous. To select only the issues whose scans are complete, see Completeness — a size threshold such as num_pages >= 12 does not do it.

Size and coverage

• 1,852 page images across 180 issues
• Date range: 2012-02-18 to 2016-12-31
• Roughly 4 GB of image data
• 1,400 PNG and 452 JPEG source files

How 183 source folders become 180 issues

The source is 183 folders holding 1,864 image files. Every file is accounted for, as follows:

	Folders	Files
Source on disk	183	1,864
Empty in source: july 5 2014, nov 16 2013	-2	-0
Excluded duplicate: august 10 2013	-1	-12
Published	180	1,852

The 12 excluded files are byte-identical (verified by SHA-256) to the contents of the separate august 8 2015 folder, are themselves named august 8 2015_page-0001..0012.jpg, and their printed masthead reads "Vol 82, No. 32 | Agost 8, 2015 | 12 Pages". They are a copy of the 2015-08-08 issue misfiled under a 2013 folder name. They are excluded because publishing them would assert the date 2013-08-10 for pages that are demonstrably from 2015-08-08. No unique image is lost: the same twelve images are published under their correct date via the august 8 2015 issue.

Whether an issue actually dated 2013-08-10 ever existed in this collection cannot be determined — that folder's contents had already been replaced in the source.

• Page images are A4-proportioned, RGB, at one of two resolutions: 1654x2339 (200 dpi) or 1241x1754 (150 dpi)

Issues and pages per year

Year	Issues	Pages
2012	42	518
2013	47	563
2014	30	253
2015	25	212
2016	36	306
Total	180	1,852

The paper is a Saturday weekly, and 175 of the 180 issues fall on a Saturday. Five do not (august 13 2013, june 24 2016, november 23 2014, october 13 2015, september 9 2012); these are most likely day-number typos in the original folder names. They have been left exactly as found rather than silently "corrected", since the folder name is the only available date evidence.

Completeness

Most issues in this dataset are partial scans, and page is a sparse label rather than a dense 1..N index. This is the single most important thing to know before training on the data.

page carries the newspaper's own printed pagination, taken from the source filenames. It is not a row index and it is not guaranteed to start at 1 or to be consecutive. Of the 180 issues:

	Issues
Complete: contiguous 1..N, front page present	32
Partial: front page missing and internal gaps	148

The two groups line up exactly with the two source-scan generations:

• The 32 complete issues (452 JPEG pages) run contiguously from page 1.
• The 148 partial issues (1,400 PNG pages) all lack page 1 — the front page — and additionally have internal gaps. They begin at page 2 (123 issues) or page 3 (25 issues).

Across the corpus 553 page slots are absent between the lowest and highest page number present. Per-issue shortfall (max(page) - num_pages):

Missing pages in issue	Issues
0 (complete)	32
2	1
3	70
4	52
5	19
6	4
7	2

Two real examples:

december 13 2014 -> pages 2, 3, 5, 8, 9, 10, 11      (7 scans; 1, 4, 6, 7 absent)
december 17 2016 -> pages 3, 4, 5, 8, 11             (5 scans; max page is 11)

This is an artifact of how the PNG-era issues were originally retrieved, not of this packaging — the missing files are simply not in the source.

What this means in practice

• Page 1 is absent from 148 of 180 issues. Any model or evaluation that depends on front pages, mastheads, or lead articles has only 32 issues to work with. Filter with page == 1.
• Do not treat page as a dense index, and do not assume row k of an issue is page k+1.
• num_pages is scans available, not issue length. Mastheads on complete issues show 12 or 16 printed pages.

The dataset does not ship a max_page column, but it is one line to derive, and comparing it with num_pages is the quickest way to find the partial issues:

import collections
from datasets import load_dataset

ds = load_dataset("Reubencf/konkani-newspapers", split="train")

# derive max_page and num_pages per issue in one pass over the metadata only
meta = ds.remove_columns("image")          # avoids decoding any images
max_page, num_pages = collections.defaultdict(int), {}
for r in meta:
    max_page[r["issue"]] = max(max_page[r["issue"]], r["page"])
    num_pages[r["issue"]] = r["num_pages"]

complete = {i for i in num_pages if num_pages[i] == max_page[i]}
print(len(complete))                        # 32

# keep only rows belonging to complete issues
ds_complete = ds.filter(lambda r: r["issue"] in complete)

# the 32 surviving front pages
front_pages = ds.filter(lambda r: r["page"] == 1)

Page-count distribution

Pages present per issue (min 5, max 16, mean 10.3):

Pages in issue	Number of issues
5	3
6	4
7	30
8	31
9	2
10	14
11	24
12	54
13	1
16	17

Caveats and known data-quality issues

Please read this section before training on the data.

See Completeness above for the most important caveat: most issues are partial scans and page is a sparse label.

Four pages are duplicated within their own issue. In jan 19 2013 (pages 5/6 and 7/11), december 29 2012 (pages 8/14), and jan 17 2015 (pages 5/7), the same image is stored under two different page numbers. One of each pair carries a wrong page number, but there is no way to tell which, so both have been kept.

Two dates are missing entirely because their source folders were empty. The folders july 5 2014 and nov 16 2013 exist in the source but contain zero files, so there are no rows for 2014-07-05 or 2013-11-16. See How 183 source folders become 180 issues for the full reconciliation, including the excluded duplicate folder.

Folder-name dates were normalised from inconsistent spellings. The source folders are named <month> <day> <year>, but the month token is not consistent: alongside the twelve full month names it includes the abbreviations jan, feb, jun, nov and two outright misspellings, decenber (for December, the decenber 7 2013 issue) and ocotber (for October, the ocotber 19 2013 issue). All of these were normalised when computing date, year, month, and day. The issue column preserves the original folder name verbatim, including the misspellings, so every row can be traced back to its source directory. If you re-derive dates yourself from issue rather than using the date column, you must handle these spellings or you will silently drop those issues.

Source filenames used five different conventions. Page numbers were parsed from all of them: page_N.png (1,400 files), <issue>_page-NNNN.jpg under August, April and December prefixes (440 files), and august 6 2016_pages-to-jpg-NNNN.jpg (12 files — note pages-to-jpg, not page-). The zero-padding also varies: page_N is unpadded, so a lexicographic sort of the source filenames puts page 10 before page 2. The page column is a parsed integer and sorts correctly.

No OCR ground truth. There are no transcriptions, bounding boxes, or article segmentations. Scan quality varies and some pages show skew, bleed-through, and compression artifacts.

Rights

These scans are of third-party newspapers. Copyright in the newspaper content is retained by the publishers of Vauraddeancho Ixtt and by the individual authors, photographers, and illustrators whose work appears in it.

The uploader of this dataset is not the rightsholder, has not obtained a licence from the publisher, and has not asserted or granted any licence over this material. The license field is deliberately set to unknown rather than to a permissive licence, because no permissive licence has been granted.

The dataset is shared to support research on Konkani, a low-resource language. Downstream users are responsible for clearing rights themselves. Before using this data — particularly for any commercial purpose, redistribution, or publication of reproductions — you should satisfy yourself that your use is lawful in your jurisdiction, whether under fair use, fair dealing, a text-and-data-mining exception, or a direct licence from the publisher. If you are the rightsholder and want this dataset amended or taken down, please open a discussion on this repository.

Citation

If you use this dataset, please credit the newspaper as the source of the underlying material:

@misc{konkani_newspapers,
  title  = {Konkani Newspapers: scanned pages of Vauraddeancho Ixtt (2012--2016)},
  author = {Reubencf},
  note   = {Scanned pages of the Konkani weekly Vauraddeancho Ixtt, Pilar, Goa.
            Copyright in the underlying material is retained by the publisher.},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Reubencf/konkani-newspapers}
}