MIZAN and CAMP: an Arabic poetry rhyme evaluation, and a proposed fix

We asked six small open language models to write Arabic poems from short personal stories, then checked the poems for rhyme and for the layout we asked for. Almost none kept to the rules.

By: Hassan Barmandah, Umm Al-Qura University

Why this matters

Classical Arabic poetry follows strict rules. The rhyme (qafiyah) sets the sound every line ends on. Language models write fluent Arabic, but the small models we tested rarely follow these rules, and nothing in their output shows that anything is wrong.

One likely reason is diacritics, the small marks that show short vowels. Rhyme depends on them, and most Arabic text online leaves them out. Common Arabic benchmarks such as ArabicMMLU and AlGhafa test knowledge and language skill, not poem structure.

What we did

We wrote 40 short first-person stories with help from an AI assistant, 4 for each of 10 genres (praise, elegy, love, longing, parting, reproach, sadness, romance, religious and patriotic). The genre labels come from arbml/ashaar, and each story was written from the title and verses of a different poem there. Every prompt asks for a poem of six bayts (a bayt is one verse line made of two halves) on one rhyme (the prompt also names one meter, which this evaluation does not score), with full diacritics and the two halves of each bayt separated by three dots (...).

Six models wrote 3 poems for every prompt, so 120 per model and 720 in total: Qwen2.5-1.5B-Instruct, Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, Gemma-2-2b-it, Jais-family-1p3b-chat and Falcon-H1-3B-Instruct. All of them used the same settings (temperature 0.8, top-p 0.9, up to 700 new tokens).

A poem counts as following the format if it has at least 3 lines that split into exactly two halves at the three dots (or *, **, #), each half with at least 2 Arabic words. Lines with Latin letters, parentheses or colons are ignored, since they are notes and not verse. A poem has one rhyme if its distinct final words all end on the same rhyme letter, read with bohour. A word repeated at the end of every line counts once, so a refrain cannot pass, and at least 3 different endings are needed. The "either layout" column applies the same rhyme check to every second line, for poems written one half per line.

We sampled 200 poems written by people from Ashaar. Of those with at least 3 bayts, 68% pass the rhyme check.

Results

Model	Followed the format	One rhyme, strict format	One rhyme, either layout
Qwen2.5-1.5B	0.0%	0.0%	0.0%
Qwen2.5-3B	3.3% [0.8, 6.7]	0.0%	1.7% [0.0, 4.2]
Gemma-2-2b	0.0%	0.0%	0.0%
Llama-3.2-3B	0.0%	0.0%	4.2% [0.8, 9.2]
Jais-1.3B	0.0%	0.0%	0.0%
Falcon-H1-3B	18.3% [11.7, 25.0]	0.8% [0.0, 2.5]	3.3% [0.8, 6.7]

• No model keeps one rhyme letter through a whole poem more than about 4% of the time, and the intervals overlap, so this evaluation cannot rank the six models. For comparison, 68% of human poems pass the same rhyme check.
• Falcon writes two halves per line in about 1 poem in 5, and Qwen-3B in 1 in 30. The other four never do. Falcon often separates the halves with * or *** instead of the three dots, and the check accepts those. Gemma and Llama use the three dots, but at the start or end of a line instead of between the two halves. Jais ignores them.
• Among the poems that did follow the format, Falcon puts 43% of its bayts on the most common rhyme letter, against 21% for Qwen-3B.
• Eleven poems pass a rhyme check. We read all of them. In five (one from Qwen-3B, four from Llama) the second halves really end on one rhyme letter, though none reads as good verse. In the other six the check was wrong: reading the endings shows two different letters (for example ي and ق), so the true rate is lower than the table says.

Example outputs

Real outputs, shortened to their first lines.

A refrain does not count as a rhyme

Qwen2.5-3B. Story: جاري حمّلني ذنبًا لم أرتكبه أمام الناس، وكنتُ وفيًّا له. أريد أن أعاتبه بلطف وأذكّره بعشرتنا.

فِي البحْرِ مَن يَحْمِلُ الذَّنْبَ أَوْ لاَ... يَعْتَبِرُ وَلاَ يُذْكَرُ أَوْ لاَ
فِي البحْرِ مَن يَفْعَلُ الْخَيْرَ وَلاَ... يُعَاتَبُ وَلاَ يُذْكَرُ أَوْ لاَ
فِي البحْرِ مَن يُؤَدِّيَ الدَّينَ وَلاَ... يُعَاتَبُ وَلاَ يُذْكَرُ أَوْ لاَ
فِي البحْرِ مَن أَمَّنَ الأَخِيْرَ وَلاَ... يُعَاتَبُ وَلاَ يُذْكَرُ أَوْ لاَ

It followed the format, but every bayt ends in the same words, so it fails the rhyme check.

Verse-like lines in another layout

Gemma-2-2b. Story: غادرتُ بلدة نجد وقلبي بقي هناك، ولا أجد بعد فراق أهلها غير البكاء والأنين.

غادرتُ بلدة نجد وقلبي بقي هناك،
... وَلَا أجد بعد فراق أهلها غير البكاء والأنين.
ملامحها في عيونِي تنعكس،
... كأنها حلمٌ في صمتِ الليل،

The three dots start the second line instead of sitting inside it, so it does not count as following the format. The rhyme check on every second line also fails.

Broken Arabic

Qwen2.5-1.5B. Story: كنتُ في شبابي أندفع وراء أهوائي، ثم تغيرتُ حين تعلمتُ الصبر والحكمة من شيخ كبير نصحني. أريد أن أمدح هذا الرجل الذي غيّر طريقي.

في سن شبابي، كنت أ奔 متعه.
أطلقت الأفكار كالرياح،
لكن مع مرور السنين،
تعلمت الصبر والحكمة،

The first line contains a stray Chinese character, and 5 of this model's 120 outputs do. The cleaning step removes Latin letters but not other scripts.

One of the few passes

Llama-3.2-3B. Story: جلستُ مع جدي يحكي لي عن أيام الأجداد وكيف كانت بلادنا قوية بهم. أتمنى أن يعود ذلك العهد على الأحفاد.

نُجيتَ بِحُبِّ الأَيّامِ الَّتي
بِفَضائِلٍ كَأَنَّهُنَّ طَعمُ
لَم يَخَطِلِ العَهدُ مِن عَشيرَةٍ
عَلَيكَ بِالعُمرِ وَبِالنِعَمِ
فَطَولُ الأَيّامِ تُجَلِّدُ
وَالعَهدُ يَعودُ عَلَيكَ في الأَيّامِ

It writes one half per line, and the second, fourth and sixth lines all end in م. It passes the either-layout rhyme check, although the lines mean little.

Limits

• The test is small: 40 prompts and 120 poems per model, and only small open models. It says nothing about larger models.
• The rhyme letter comes from bohour, a rule-based tool. Beyond the test on human poems, we did not check its choices by hand.
• Outputs often repeat themselves or contain stray characters from other scripts, such as Chinese.

What is included

The dataset has the 40 prompts with their stories, genres, and source poem titles. It also has the 720 poems with their scores, a results table and the chart. In the scores file, a rhyme share of 0 and a rhyme result of false mean the poem had no two-half lines (or too few lines), so there was no rhyme to measure.

Trailing blank lines were removed from the poems, because Gemma pads many outputs with hundreds of them.

The notebook that produced the results is included. Generation ran earlier on a GPU, so the generation cell reports "skip (already done)" and keeps its original output. The scoring cells were re-run with the final scoring rules. The research proposal is in the proposal folder.

Proposed path forward

Read the full proposal: proposal/Research_Proposal.pdf (4 pages, click to open).

The plan has two parts.

1. Fix the poetry gap. Fine-tune a small Arabic-native model with LoRA+ on story and poem pairs built from Ashaar, keeping the diacritics. When it writes, have it produce a few versions and keep only the ones the rule-based tools in this evaluation accept.
2. Add the skill without forgetting. CAMP (Consolidated Adaptive Merge for PEFT) is our method for adapting a model to a new task without losing what it knew. It probes how much the new task conflicts with each layer, gives conflicting layers more adapter capacity, and merges the new adapter into the previous one.

CAMP is an early prototype, tested only on a small three-layer network with synthetic tasks, not on a real language model or on poetry. None of this has been trained on poetry yet, so it remains a plan.