Chandohasam: Telugu metrical poetry corpora

Four corpora of Telugu metrical poems (padyams) in one record format, with metre labels, word glosses and meanings in Telugu and English. They are the evaluation corpora of Chandohasam, a project on generating Telugu metrical poetry under exact metrical constraints; the project's 120M masked diffusion language model is at samvaran/chandohasam.

15,415 records, 12,704 of them verse poems (59,147 verse lines in 28 metres); the rest are prose passages of the Bhāgavatamu.

File	Work	Records (verse)	Metre labels	Glosses and Telugu meanings	English meanings
bhagavatam.json	Bammera Pōtana, Āndhra Mahābhāgavatamu (15th century)	10,066 (7,366)	from the source compilation; 27 corrected	from the source compilation (9,018 records)	machine translation (9,018)
vemana.json	Vēmana's śatakam	1,164 (1,164)	heuristic (1,022), metre engine (142)	machine-written	machine translation
kuchimanchi_timmakavi.json	Kūcimañci Timmakavi's pure-Telugu Rāmāyaṇamu	1,653 (1,642)	metre engine (1,555), Chandassu (53), partial (3), none (42)	machine-written	machine translation
chandassu.json	25 śatakams of the Chandassu dataset	2,532 (2,532)	from the Chandassu dataset	machine-written	machine translation

No poem appears in two files: Chandassu poems already present in the Vemana or Kuchimanchi files were dropped when it was imported.

Loading

from datasets import load_dataset

bhagavatam = load_dataset("samvaran/chandohasam", "bhagavatam", split="train")
print(bhagavatam[0]["verse"], bhagavatam[0]["bhavam_en"])

Each file is a JSON array of records, so it can also be read with json.load.

Record format

Every file uses the same fields; a field a corpus does not have is null or empty.

Field	Meaning
id	Unique id. Bhāgavatamu: the source compilation's scheme <skandha>-<number>-<metre code>., e.g. 1-1-శా.; Vemana vemana-<n>-<code>.; Kuchimanchi kuchimanchi-<n>; Chandassu chandassu-<śatakam>-<n>-<code>.
skandha	Bhāgavatamu: the book (1–12, with 5.1/5.2 and 10.1/10.2); Chandassu: the śatakam
poem_number, sub_number	Position in the work; sub_number > 0 marks a record that continues another, such as the poem that closes a sīsa
parent_id, child_id	Bhāgavatamu: each of the 1,048 sīsa poems is linked to the record that closes it (a tēṭagīti or āṭaveladi, rarely another record)
metre_code, metre, metre_roman	Metre: abbreviation, Telugu name, romanised name (kandamu, aataveladi, seesamu, …); for prose, vachanamu, gadya or dandakamu
metre_variant	Two Bhāgavatamu records: a sīsa with its closing ettugīti, and a five-line poem
form	verse or prose
expected_lines, allowed_lines	Lines the metre expects, and the counts it accepts
line_count, complete	Lines present; whether the poem is complete (117 Chandassu poems are not)
verse	The lines (pādas, or the printed half-lines of a sīsa)
teeka, teeka_pairs	Word-by-word gloss (ṭīkā), as text and as [{word, meaning}]; Bhāgavatamu only
bhavam	Telugu meaning (bhāvamu); Bhāgavatamu only
bhavam_en	English meaning: a machine translation of the Telugu meaning (bhavam, or generated.bhavam)
generated	Machine annotation of the other three files: {run, model, rescued, bhavam, prathipadartham: [{word, split, meaning}], …}
source	Provenance of the record. Bhāgavatamu: original (8,207), lines_restored (1,263), inserted (586), user-supplied (10); the others: the source file
author	Poet (Vemana, Kuchimanchi files)
label_source	How the metre label was obtained: heuristic, engine, engine-partial, chandassu, corpus
metre_corrected, verse_corrected	On corrected records only: the original label or lines, with the date and the reason

How the files were made

• Bhāgavatamu. The verse, word glosses and Telugu meanings of the compilation Pōtana Telugu Bhāgavatamu by Ūlapalli Sāmbaśiva Rāvu (telugubhagavatam.org), which is also transcribed on Telugu Wikisource. The compilation names no single printed edition. A sīsa and the poem that closes it are separate records; the meaning of the whole poem is on the closing record.
• Vemana and Kuchimanchi. Poem texts converted into this record format. Vemana's labels come from a heuristic (four-line poems are āṭaveladi, eight-line ones sīsa), with 142 corrected by the project's metre engine; Kuchimanchi's labels were assigned by the metre engine.
• Chandassu. The Kaggle dataset boddusripavan111/chandassu by Boddu Sri Pavan and Boddu Swathi Sree (4,651 rows from 28 śatakams, texts from andhrabharati.com). Its rows give 3,979 poems; the 1,447 already in the Vemana or Kuchimanchi files were dropped, leaving 2,532.
• Machine annotation. For Vemana, Kuchimanchi and Chandassu, the word-by-word glosses and Telugu meanings in generated were written by Gemini (gemini-3.5-flash-lite; 59 answers come from gemini-3.8-flash, used when the first answer failed the pipeline's automatic checks). Every English meaning (bhavam_en) is a Gemini translation of the Telugu meaning. Letters of other scripts that leaked into the machine annotation were removed, partly by rule and partly by hand with an AI assistant; each change is kept in generated.script_fixes. Nobody has reviewed the machine annotation.
• Corrections. Metre labels and verse lines found wrong by the project's metre engine and data audit were corrected in place; every corrected record keeps the original in metre_corrected or verse_corrected.

Quality

Share of the verse poems whose gaṇa (syllable weights), prāsa and yati all hold under the project's metre engine, with the relaxed (strict) rule profile:

File	Valid, relaxed (strict)
Bhāgavatamu	91.2% (88.7%)
Vemana	83.5% (82.6%)
Kuchimanchi	90.6% (88.7%)
Chandassu	74.0% (69.9%)

A poem that fails may be a genuine irregularity, a transcription error or a reading the engine does not accept (yati, in particular, depends on sandhi that the script does not show).

Considerations

• Refrains repeat. 892 of the 1,164 Vemana poems end with the same refrain line (makuṭa), విశ్వదాభిరామ వినర వేమ., and 77.1% of Vemana poems share a whole pāda with another poem (Chandassu 25.3%, Kuchimanchi 11.8%, Bhāgavatamu 1.1%). Keep poems that share a refrain in the same split.
• Machine-written meanings are unreviewed and can be wrong, especially for archaic words and long compounds.
• The Bhāgavatamu meanings sometimes run on into commentary (notes on prāsa or figures of speech); the English translation keeps only the meaning.

License and attribution

This dataset is released under CC BY-SA 4.0. It contains:

• the poems of Bammera Pōtana, Vēmana and Kūcimañci Timmakavi, which are in the public domain;
• the word glosses and Telugu meanings of the Āndhra Mahābhāgavatamu from Ūlapalli Sāmbaśiva Rāvu's compilation Pōtana Telugu Bhāgavatamu (telugubhagavatam.org), as transcribed on Telugu Wikisource, whose text is under CC BY-SA;
• poems from the Chandassu dataset by Boddu Sri Pavan and Boddu Swathi Sree, released on Kaggle under the MIT License;
• machine-written glosses, meanings and translations produced for this project.

When you use the data, credit these sources. If you use the Chandassu poems, please also cite:

@misc{pavan2025computationalsociallinguisticstelugu,
  title  = {Computational Social Linguistics for Telugu Cultural Preservation: Novel Algorithms for Chandassu Metrical Pattern Recognition},
  author = {Boddu Sri Pavan and Boddu Swathi Sree},
  year   = {2025},
  eprint = {2510.01233}
}

Citation

@misc{chandohasam_corpora_2026,
  title        = {Chandohasam: Telugu metrical poetry corpora},
  author       = {Rallabandi, Samvaran Kashyap and Emani, Mahesh and Salopanthula, Radhe Shyam},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/samvaran/chandohasam}}
}