---
license: apache-2.0
language:
- ar
base_model: unsloth/Qwen3-8B
tags:
- arabic
- poetry
- qwen3
- gguf
- lora
- experimental
pipeline_tag: text-generation
---

# Arabic Poetry Qwen3-8B (LoRA, GGUF) — Experimental

> **Status: experimental / hobby project.** This model produces *coherent
> Arabic in a loose poetic register*. It does **not** reliably produce correct
> classical meter (بحر) or rhyme (قافية), and output quality is uneven. Set
> expectations accordingly. Read the Limitations section before using.

A LoRA fine-tune of **Qwen3-8B** on a corpus of Arabic poetry that was OCR'd
locally from 19 books. Trained as continued-pretraining on discrete poems.

## What it actually does

- Generates Arabic text with a poem-like shape (short lines, stops cleanly)
- Stays in Arabic, on the prompt's theme more than the base model does
- Style leans modern/free-verse (Darwish-ish), not classical ode

## What it does NOT do well

- **No reliable meter or rhyme.** It does not scan to a specific بحر.
- Imagery is often weak or vague; some lines are semantically loose.
- Quality varies a lot run-to-run.
- Classical/Jahiliyya register is weak (the OCR corpus had artifacts).

This is a **data-and-scale-limited** result: ~2,244 OCR'd poems (with residual
OCR noise) on an 8B model is not enough to install genuine Arabic prosody. It
is shared as an experiment and a starting point, not a finished poetry engine.

## Usage (LM Studio / llama.cpp)

Download `arabic-poetry-qwen3-8b-f16.gguf`. Recommended sampling:

- Temperature `0.7`, Top-p `0.92`, Top-k `40`
- **Repeat penalty `1.3`** (lower values loop)
- Reasoning/thinking: **OFF**
- System prompt: empty

Prompt with Arabic openers, not English instructions:

```
قصيدة في وصف الصحراء:
في حضرة الغياب،
أحبكِ يا وطني،
```

> Note: send prompts via a UTF-8-correct client. Some terminal/curl setups on
> Windows mangle Arabic UTF-8 and will make the model emit garbage — that is a
> client encoding bug, not the model.

## Training

| | |
|---|---|
| Base | `unsloth/Qwen3-8B` (full bf16, not quantized) |
| Method | LoRA r=32 α=32, attention-only, dropout 0.05 |
| Data | 2,244 discrete cleaned Arabic poems, EOS-terminated |
| Schedule | 2 epochs, cosine LR 1.2e-4, manual training loop |
| Selection | best pre-overfit checkpoint by sample quality (not final) |
| Hardware | single RTX 4090, ~19 min |

## License

Apache-2.0 (inherits from Qwen3-8B).

## Honest note

Built end-to-end on a single workstation (OCR → corpus cleaning → LoRA → GGUF).
The most reusable artifact from the project is arguably the cleaned corpus and
pipeline, not this particular adapter. Contributions / a cleaner meter-labelled
Arabic corpus would meaningfully improve a v2.