Datasetpublic

artindnr/poem-generation-1k

poem-generation-1k poem-generation-1k is a synthetic dataset of 1,149 Persian poem-request instructions paired with generated poems, spanning 11 classical and modern Persian poetic forms and 35 topic categories. It is the base dataset in the Poem Generation Collection, and forms the source pairs for artindnr/poem-reasoning-1k, which adds reasoning traces on top of these same (question, poem) pairs. How it was built Both the poem-request instructions and the poems… See the full description on the dataset page: https://huggingface.co/datasets/artindnr/poem-generation-1k.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

3likes74downloads

Dataset Card

poem-generation-1k

poem-generation-1k is a synthetic dataset of 1,149 Persian poem-request instructions paired with generated poems, spanning 11 classical and modern Persian poetic forms and 35 topic categories. It is the base dataset in the Poem Generation Collection, and forms the source pairs for `artindnr/poem-reasoning-1k`, which adds reasoning traces on top of these same (question, poem) pairs.

How it was built

Both the poem-request instructions and the poems themselves were synthetically generated using the GPT-5.6 series of models. For each example, a poem type and topic were combined into a natural Persian instruction (e.g. "یک دوبیتی در حال‌وهوای امید به آینده دربارهٔ جوانه‌ای که از دل خاک سر برمی‌آورد بسرای" — "write a دوبیتی in the mood of hope for the future, about a sprout breaking through the earth"), and a poem matching the requested form, mood, and imagery was generated in response.

Dataset structure

field	type	description
`id`	int64	Row identifier
`poem_type`	string	Persian poetic form requested (11 classes — see below)
`topic_category`	string	Topic/mood the poem should evoke (35 classes — see below)
`messages`	list	Chat-formatted example: a `system` turn ("تو یک مدل هوش مصنوعی شاعر هستی." — "You are an AI poet model."), a `user` turn (the Persian poem request), and an `assistant` turn (the generated poem)

—Rows: 1,149
—Splits: train (single split)
—Format: Parquet

Poem types (`poem_type`)

Includes دوبیتی, رباعی, غزل, مثنوی, قصیده, قطعه, ترجیع‌بند, ترکیب‌بند, چهارپاره, هایکوی فارسی (Persian haiku), and شعر نو (نیمایی) (Nima-style free verse) — 11 forms spanning classical fixed-meter forms through modern free verse.

Topic categories (`topic_category`)

35 categories covering everyday life and nature themes (زندگی روزمره, بهار, پاییز, زمستان, باران, کوهستان, دریا, ماه, ستاره و آسمان), emotional/relational themes (عشق, دوستی, مادر, تنهایی, فراق و دوری, غم و اندوه, دلتنگی برای خانه, غربت, وطن), and reflective/spiritual themes (امید, امید به آینده, مرگ و فنا, عرفان و سلوک, می و شراب عرفانی, زمان و گذر عمر, آزادی, صلح), among others.

Intended use

Built for supervised fine-tuning (SFT) of Persian language models on poem generation across a range of classical and modern forms. Can be used standalone as a non-reasoning baseline, or paired with poem-reasoning-1k for training with reasoning traces on the same underlying pairs.

Limitations

—All content — both instructions and poems — is synthetically generated; poems have not been manually reviewed for classical prosodic correctness (وزن/قافیه) at scale.
—Poem type and topic coverage is not perfectly balanced across the 11 × 35 combinations.

Citation

If you use this dataset, please cite:

@misc{poem-generation-1k,
  author = {artindnr},
  title = {poem-generation-1k: Synthetic Persian Poem Generation Dataset},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/artindnr/poem-generation-1k}}
}