MBZUAI/instructpoet-ar

---
pretty_name: Arabic Poetry IFT
language:
- ar
language_creators:
- found
license: other
multilinguality: multilingual
size_categories:
- 1M<n<10M
source_datasets:
- original
task_categories:
- text-generation
- question-answering
task_ids:
- text2text-generation
- multiple-choice-qa
annotations_creators:
- expert-generated
- machine-generated
tags:
- arabic
- poetry
- instruction-following
- dialects
- literary-text
configs:
- config_name: generation_templates
default: true
data_files:
- split: generation_templates
path: "Templates - Poetry Generation.csv"
- config_name: continuation_templates
data_files:
- split: continuation_templates
path: "Templates - Poetry Continuation.csv"
- config_name: analysis_templates
data_files:
- split: analysis_templates
path: "Templates - Poetry Analysis.csv"
- config_name: corruption_templates
data_files:
- split: corruption_templates
path: "Templates - Poetry Corruption.csv"
---

# Arabic Poetry IFT

## Dataset Summary

Arabic Poetry IFT is a large-scale instruction-following dataset for Arabic poetry understanding and co-creation. It supports four task families: generation, continuation, revision/restoration, and multiple-choice analysis. The dataset covers Modern Standard Arabic (MSA) and four regional Arabic varieties used in the instruction layer: Gulf, Levantine, Nile Valley, and North African Arabic.

This release accompanies the ACL 2026 paper *Instruction-Guided Poetry Generation in Arabic and Its Dialects*.

**Dataset curators:** Abdelrahman Sadallah, Kareem Elozeiri, Mervat Abassy, Rania Elbadry, Mohamed Anwar, Abed Alhakim Freihat, Preslav Nakov, Fajri Koto.

### Supported Tasks

1. `generation`: compose a poem from constraints such as title, poet, era, genre, meter, rhyme, keywords, or key phrases
2. `continuation`: continue a partial poem while preserving poetic constraints
3. `revision` / `corruption`: restore a corrupted poem to its intended form
4. `analysis`: answer multiple-choice questions about poem metadata such as poet, title, keywords, meter, era, genre, and rhyme

### Language Coverage

- Modern Standard Arabic (MSA)
- Gulf Arabic
- Levantine Arabic
- Nile Valley Arabic
- North African Arabic

Most source poems are in MSA, while dialectal coverage is introduced primarily through manually written instruction templates and a smaller amount of dialectal source material.

## Dataset Construction

### Source Collection

The underlying poetry corpus was aggregated from public literary resources, unified into a common format, enriched with metadata, and deduplicated before instruction generation. Poems with only one verse were removed. Rhyme was automatically inferred when at least 70% of verse endings matched after normalization.

Training corpus statistics from the paper:

| Source | Train poems | Avg. verses |
|---|---:|---:|
| Ashaar | 123,581 | 19.81 |
| PoetsGate* | 112,482 | 15.58 |
| Adab* | 70,277 | 35.33 |
| AraPoems | 62,963 | 22.01 |
| Diwan* | 38,005 | 22.65 |
| Mawsooaa* | 18,002 | 10.25 |
| Arapoet* | 1,303 | 9.25 |
| Arabic Poetry Dataset | 662 | 19.41 |
| Arabic-Poetry-Melody | 48 | 21.44 |
| Adab World* | 6 | 93.33 |
| Other | 8 | 24.88 |
| **Total** | **427,337** | **21.39** |

Test benchmark statistics from the paper:

| Source | Test poems | Avg. verses |
|---|---:|---:|
| FannOrFlop (Al Ghallabi et al., 2025) | 6,984 | 17.97 |

\* Scraped sources, as reported in the paper.

### Metadata Enrichment

The corpus uses both original and derived metadata. Key fields used to build the instruction tasks include:

- `poem_text`
- `poem_title`
- `poet_name`
- `poet_era`
- `genre`
- `meter`
- `rhyme`
- `keywords`
- `key_phrases`

`keywords` and `key_phrases` were automatically generated using Gemini 2.5 Pro. The paper reports a manual quality check on 100 sampled poems, where 96% of extracted keywords were judged to be good quality.

### De-duplication and Leakage Control

- Intra-source duplicates were removed after normalization.
- Orthographic normalization included removing elongation, removing diacritics, and standardizing spelling variants.
- Any poem overlapping with the FannOrFlop benchmark was removed from training to avoid train/test leakage.

### Instruction Template Creation

The instruction layer was manually designed for four task families and then expanded across five Arabic varieties. The camera-ready paper reports:

| Task family | Base templates | Arabic varieties | Dialect-specific templates |
|---|---:|---:|---:|
| Generation | 246 | 5 | 1,230 |
| Continuation | 176 | 5 | 880 |
| Analysis | 214 | 5 | 1,070 |
| Revision | 8 | 5 | 40 |
| **Total** | **644** | **5** | **3,220** |

Dialect templates were written and revised by native speakers of the corresponding regional varieties.

## Dataset Statistics

### Overall IFT Size

The instruction-following dataset contains **1,350,897** training pairs and **24,815** test pairs, for **1,375,712** total examples.

| Task | Train examples | Test examples | Train subtasks | Test subtasks |
|---|---:|---:|---:|---:|
| Generation | 427,337 | 6,984 | 19 | 19 |
| Continuation | 427,276 | 6,984 | 11 | 11 |
| Revision | 68,947 | 3,863 | 8 | 8 |
| Analysis | 427,337 | 6,984 | 16 | 14 |

### Task-Specific Notes

- Continuation examples are created by splitting poems at random cut points between 10% and 90% of the poem length.
- Analysis is framed as multiple-choice question answering with 1 correct answer and 4 distractors.
- Revision examples are generated from automatically corrupted poems paired with their clean originals.

The eight revision corruption types are:

- `rhyme_structure`
- `full_style`
- `rhyme_substitution`
- `rhyme_content`
- `era_corruption`
- `meter_transformation`
- `meter_destruction`
- `meter_inconsistency`

### Metadata Distribution Highlights

Top values reported in the paper:

- Meter: Al-Tawil 20.31%, Al-Kamil 16.10%, Al-Basit 12.27%
- Poet era: Modern 37.17%, Abbasid 22.59%, Mamluk 10.05%
- Genre: General 22.89%, Short 11.61%, Praise 8.01%

## Dataset Structure

### Release Organization

The Hugging Face dataset viewer is configured with four subsets so each template file appears separately in the subset dropdown:

- `generation_templates`
- `continuation_templates`
- `analysis_templates`
- `corruption_templates`

These subsets map to the following files:

- `generation_templates` -> `Templates - Poetry Generation.csv`
- `continuation_templates` -> `Templates - Poetry Continuation.csv`
- `analysis_templates` -> `Templates - Poetry Analysis.csv`
- `corruption_templates` -> `Templates - Poetry Corruption.csv`

At minimum, each example contains the columns provided in its source CSV file.

### Example Schema

Generation / continuation / analysis template files include columns such as:

- `Placeholder`
- `Output` or `output`
- `MSA`
- `Nile Valley`
- `North Africa`
- `Gulf`
- `Levant`

The corruption template file includes columns such as:

- `Placeholder`
- `corruption_type`
- `MSA`
- `Nile Valley`
- `North Africa`
- `Gulf`
- `Levant`

## Intended Uses

### Direct Use

- Instruction tuning for Arabic poetry-capable language models
- Evaluation of controllable Arabic poetry generation
- Study of dialectal prompt robustness in Arabic
- Research on meter-, rhyme-, style-, and metadata-conditioned generation
- Benchmarking poetry analysis with multiple-choice supervision