AlbertoChestnut/telugu-ocr

Telugu OCR Dataset

A corpus of aligned scanned page images and human-transcribed Telugu text, sourced from Telugu Wikisource. Built for OCR model training and evaluation.

Stats

Total page pairs	~25,565
Books	221
Total size	~11 GB
License	CC BY-SA 4.0

Dataset Structure

dataset/
  <book_title>/
    page_0001.jpg   ← scan image
    page_0001.txt   ← transcribed Telugu text (UTF-8)
    page_0004.jpg
    page_0004.txt
    ...
    meta.json       ← per-book metadata
corpus_profile.json ← index of all books with quality and curation details
curation_decisions.json ← approved/excluded status for every book in the manifest

Page numbers are not always consecutive — gaps indicate pages with no transcription available.

---

corpus_profile.json

Top-level file. One JSON object per book covering every book in the original manifest (including excluded ones).

Field	Type	Description
index_title	string	Wikisource index page title (unique identifier)
filename	string	Original PDF filename on Wikimedia Commons
page_count	int	Total scan pages in the source PDF
mime_type	string	MIME type of source file (usually application/pdf)
size_mb	float	Source file size in MB
telugu_ratio	float	Fraction of non-whitespace, non-digit characters in the Telugu Unicode block (ఀ–౿), sampled from 3 evenly-spaced pages
telugu_pct	float	telugu_ratio as a percentage
no_transcriptions	bool	true if no transcribed pages were found on Wikisource
quality_samples	int[]	ProofreadPage quality levels from sampled pages (see Quality Levels below)
quality_avg	float	Mean of quality_samples
quality_label	string	Human-readable label for round(quality_avg)
decision	string	"approved" or "excluded"
exclusion_reason	string	Why the book was excluded (empty if approved)
decided_at	string	ISO 8601 timestamp of the curation decision
auto_decision	bool	true if decided automatically by the pipeline
profiled_at	string	ISO 8601 timestamp of language profiling

Approved books pass all three criteria: ≥ 80% Telugu script ratio, ≥ 50 scan pages, and at least one transcribed page.

---

dataset/<book>/meta.json

One file per book directory. Describes what was actually downloaded.

Field	Type	Description
status	string	"done" if completed successfully
saved	int	Number of page pairs (image + text) saved
total_pages	int	Total scan pages in the source PDF
filename	string	Original PDF filename
pages	object	Map of "page_number" → { "quality": int } for every downloaded page

The pages object contains only pages that have both an image and a transcription. Pages missing from this map either had no transcription or failed to download.

---

Quality Levels

From the Wikisource ProofreadPage extension:

Level	Label	Meaning
0	no_text	Blank or image-only page
1	not_proofread	Transcription exists but has not been reviewed
2	problematic	Flagged as having errors
3	human_proofread	Reviewed and corrected by a human editor
4	validated	Reviewed and validated by a second human editor

Levels 3 and 4 are the most reliable ground truth for OCR evaluation.

---

Source & License

All text transcriptions and original book scans are from Telugu Wikisource and are licensed under CC BY-SA 4.0. Attribution: Telugu Wikisource contributors.

---

Author

Alberto J. Chestnut Contact: data@chest.nu

---

Citation

If you use this dataset, please cite:

@dataset{chestnut2026telugu,
  author    = {Chestnut, Alberto J.},
  title     = {Telugu OCR Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  note      = {Sourced from Telugu Wikisource (CC BY-SA 4.0)},
  url       = {https://huggingface.co/datasets/AlbertoChestnut/telugu-ocr},
  doi       = { 10.57967/hf/9079 }

}