Shaer-AI/ashaar-v1-base-form

Ashaar V1 – Arabic Poetry with Meter Control

Dataset Summary

Ashaar V1 is a cleaned and structured Arabic poetry dataset designed for meter-conditioned poetry generation.

The dataset is derived from the original Ashaar corpus and has been carefully filtered, normalized, and restructured to support explicit control over Arabic poetic meter using two fields:

• base_meter — the classical Arabic meter (البحر)
• form — the structural form of the meter (تام، مجزوء، مخلع، أحذ)

This version is intended for:

• supervised fine-tuning (SFT) of large language models
• reinforcement learning with preference optimization (e.g. GRPO)
• meter-aware generation and evaluation pipelines

---

Motivation

Arabic poetic meter (ʿIlm al-ʿArūḍ) is highly structured.
Training models directly on raw labels leads to noisy or unlearnable control signals.

This dataset was built with the following principles:

• keep only learnable meter/form combinations
• remove extremely rare or noisy variants
• preserve poetic content while stabilizing control labels
• enable reliable generation of multiple candidates per meter for ranking and evaluation

---

Dataset Structure

The dataset is provided as a DatasetDict with two splits:

• train
• validation (stratified across meter and form)

Columns

Column name	Description
id	Stable numeric identifier
poem verses	List of verses composing the poem
base_meter	Classical Arabic meter (e.g. الطويل، الكامل، البسيط…)
form	Meter form (تام، مجزوء، مخلع، أحذ)
poem theme	Thematic label (when available)
poem meter	Original raw meter label from source
poem url	Source URL (for reference/debugging)

---

Supported Meters and Forms (V1)

Base meters (14)

الطويل، الكامل، البسيط، الوافر، الخفيف،
السريع، الرجز، الرمل، المتقارب، المجتث،
المنسرح، المديد، الهزج، المتدارك

Two extremely rare meters (المضارع، المقتضب) were excluded in V1 due to insufficient data.

Forms

• تام (full form)
• مجزوء (truncated form)
• مخلع (only for البسيط)
• أحذ (only for الكامل)

All other rare or inconsistent forms (e.g. مشطور، منهوك، مربع، تفعيلة) were removed to avoid noisy control signals.

---

Validation Split

The validation set is stratified by (base_meter, form) using the rule:

take = min(50, ceil(5% of each (meter, form) group))

This ensures:

• coverage of all supported meter/form combinations
• no over-removal from small meters
• meaningful evaluation of control fidelity

---

Intended Use

This dataset is intended for:

• Meter-conditioned Arabic poetry generation
• Supervised fine-tuning of LLMs (e.g. QLoRA)
• Reinforcement learning with reward models or judges
• Meter classification and verification research
• Human-in-the-loop poetic evaluation

It is not intended as a general-purpose Arabic text corpus.

---

Limitations

• Some meters have significantly more data than others
• Rare classical forms are intentionally excluded in V1
• Meter correctness depends on source annotations (though spot-checked using meter classifiers)

Future versions may introduce:

• additional meters
• curriculum-style rare forms
• rhyme annotations
• era/style conditioning

---

Citation

If you use this dataset, please cite or reference:

Ashaar V1 – Arabic Poetry with Meter Control
Shaer-AI

---

Acknowledgements

This dataset builds upon the original Ashaar poetry corpus and benefits from classical Arabic prosody scholarship and open-source meter classification models.

---

Contact

For questions, improvements, or collaborations, please contact:

Shaer-AI
https://huggingface.co/Shaer-AI