qwen3-8b-arabic-poetry-v3

qwen3-8b-arabic-poetry-v3

Experimental Arabic poetry fine-tune based on Qwen/Qwen3-8B, exported as GGUF for LM Studio and llama.cpp.

This build was trained from a local OCR-cleaned Arabic poetry corpus and merged into the Qwen3-8B base model. The goal is better Arabic poetic style, continuation, and short literary analysis.

File

SHA256:

B8C457A9CC0527283315BE465CAE42A65F92234EAC366F136FC7A7BC67B62E1B

Recommended LM Studio Settings

• GPU offload: max/all layers
• Context: 4096
• Reasoning / Thinking: OFF
• Temperature: 0.6
• Top P: 0.9
• Repeat penalty: 1.1 to 1.15
• Max tokens: 200-400

Thinking mode is not recommended for this first public test build. Local tests showed non-thinking mode stops cleanly, while thinking mode can continue until the token cap.

Hardware Compatibility

This GGUF is a Q4_K_M quantization of an 8B model, so it is intended for local inference on consumer GPUs, Apple Silicon, or CPU through LM Studio, llama.cpp, Ollama-compatible GGUF loaders, or similar runtimes.

Minimum practical memory guidance:

• Model file size: ~4.68 GiB.
• Recommended VRAM for smooth GPU use: 8 GB or higher.
• Recommended system RAM for CPU/offloaded use: 16 GB or higher.
• Start with 4096 context before trying larger context sizes.

Example Prompts

اكتب أربعة أبيات عربية فصيحة عن الحنين إلى الوطن، ولا تشرح.

حلل هذه الصورة الشعرية باختصار: الليل يسند رأسه إلى كتف النهر.

أكمل بيتين بأسلوب عربي شعري: في الصحراء نام الضوء فوق رمالها

Local Smoke Test

Tested locally with llama.cpp on RTX 4090:

• Prompt eval: ~451 tok/s
• Generation: ~121 tok/s
• VRAM model/context/compute at 4096 context: ~5.3 GB

Training Notes

• Base: Qwen/Qwen3-8B
• Fine-tune type: LoRA/QLoRA-style adapter, then merged into base
• Adapter target modules: q_proj, k_proj, v_proj, o_proj
• LoRA rank: 16
• LoRA alpha: 16
• Final validation-loss estimate from local run: ~1.184

Limitations

This is an experimental model, not a final production Arabic poetry model.

• Poetry quality is still uneven.
• OCR noise may have influenced parts of the style.
• Best used for creative Arabic poetry experiments, not factual answers.
• Use reasoning/thinking off for now.

License

The base model Qwen/Qwen3-8B is Apache-2.0. This derivative GGUF is released under Apache-2.0.