---
license: apache-2.0
language:
- ar
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
- arabic
- poetry
- shi3r
- aruud
- arabic-llm
- reasoning
- chain-of-thought
- qwen2.5
- gguf
- lora
pipeline_tag: text-generation
---

<div align="center">

# 🪶 المُحنّك · Almohanek 1.0 — 7B

### نموذجٌ عربيٌّ **يُفكِّر بالعربية في العَروض** قبل أن يَنظِم الشِّعر

**The Arabic poet that reasons in Arabic — out loud — before it writes a line.**

`by` **CyberQ** · Apache-2.0 · runs locally · 100+ tok/s on a single GPU

</div>

---

## ✨ لماذا «المحنّك»؟ / Why Almohanek?

معظم النماذج تكتب شعراً عربياً ثم تُفكّر بالإنجليزية (إن فكّرت). **المحنّك**
مختلف: يفتح وسم `<think>` ويُحلّل **البحر والقافية والصورة** *بالعربية الفصحى*،
ثم ينظم البيت — كما يفعل شاعرٌ مُتمكِّن.

Most "Arabic" models think in English behind the scenes. **Almohanek thinks in
Arabic, on the page** — it opens a `<think>` block, reasons about the **بحر
(meter), قافية (rhyme), and imagery** in fluent Arabic, *then* composes. You see
the craft, not just the output.

| | |
|---|---|
| 🧠 **Arabic chain-of-thought** | Reasons about عَروض in Arabic, never leaks English (strict-gate verified) |
| 🎓 **Distilled, not templated** | Reasoning distilled by rejection-sampling from a 32B teacher, **verified against 1.8M labelled classical verses** |
| 🪪 **Owns its identity** | "أنا المحنّك من CyberQ" — baked in, no foreign-model leakage |
| ⚡ **Local & fast** | GGUF, multiple quants, ~100+ tok/s on one consumer GPU |
| 🔓 **Open** | Apache-2.0, built on Qwen2.5-7B-Instruct |

---

## 🚀 Quickstart (LM Studio / llama.cpp)

| File | Size | Pick this if… |
|---|---|---|
| **`Almohanek1.0-7B-Q6_K.gguf`** | ~6 GB | you want the **sharpest** output ⭐ |
| `Almohanek1.0-7B-Q4_K_M.gguf` | ~4.5 GB | you're tight on VRAM |

**System prompt** (copy as-is):
```
أنت المحنّك من CyberQ، شاعر عربي خبير بالعروض. قبل أي قصيدة فكّر بإيجاز
بالعربية داخل وسم <think> ... </think> (البحر، القافية، الصورة)، ثم اكتب الشعر.
لا تذكر هويتك إلا إذا سُئلت "من أنت".
```
**Sampling:** `temperature 0.6 · top_p 0.9 · top_k 20 · repeat_penalty 1.15`

**Try it:**
```
اكتب قصيدة من ٦ أبيات على بحر الكامل في الحنين
ما بحر هذا البيت: قِفا نَبكِ مِن ذِكرى حَبيبٍ وَمَنزِلِ
```

---

## 🛠️ How it was built

Qwen2.5-7B-Instruct → QLoRA, trained with a **manual loop** and
**validation-gated** checkpoints. The reasoning was **distilled** (not hand-
templated): a 32B teacher generated Arabic عَروض rationales that were **kept only
if they were all-Arabic, bounded, and matched the ground-truth meter** of 1.8M
labelled verses. A dedicated identity pass makes it consistently introduce
itself as **المحنّك من CyberQ** without breaking its Arabic reasoning.

---

## 🧭 Honest status & roadmap

**Almohanek 1.0 is an experimental v1 — and we say so plainly.**

✅ **Solid today:** Arabic-only `<think>` reasoning · stable identity · clean
fluent Arabic · reliable stop · valid GGUF.

🚧 **On the roadmap (v2):** poetic *quality* is good-not-masterful — output is
grammatically sound Arabic but can be uneven, and meter-correctness isn't yet
guaranteed at generation time. v2 targets this directly with an **objective
عَروض quality gate** and a **curated master-poet corpus** (المتنبي, شوقي,
الشريف الرضي, البارودي, ابن زيدون …) on a larger base. This is a 7B specialist,
not a general assistant.

We ship honestly: real strengths up front, limits named, roadmap public.

---

<div align="center">

**المحنّك · Almohanek** — *حيث يلتقي العَروض بالذكاء الاصطناعي*
*where Arabic prosody meets AI* — **CyberQ** · Apache-2.0

</div>