Elgizawy-Translation-Style is a specialized fine-tune of Qwen-3-8B designed for high-quality, atmospheric, and cinematic subtitle translation.

Unlike standard translators (like Google Translate) that focus on literal meaning, this model is optimized to maintain the "flavor," tone, and dramatic weight of English sci-fi and drama scripts when converting them to Arabic.

🛠 Training & Quantization Pipeline

This model was developed using a high-precision pipeline to ensure the best possible "translation taste":

Full-Precision Training: Fine-tuned using LoRA on the full 16-bit (BF16) version of Qwen-3-8B to preserve maximum linguistic nuance.
MLX Quantization: The fine-tuned weights were then quantized to 4-bit specifically for the MLX framework to support high-performance inference on Apple Silicon.
GGUF Conversion: Finally, the model was converted to the GGUF (Q4_K_M) format using llama.cpp, making it compatible with universal tools like LM Studio and Ollama.

[!IMPORTANT] Optimized for Dialogue: This model is specifically trained for short, punchy sentences typical of subtitles and cinematic scripts. It excels at capturing emotional nuance in brief exchanges. For best results, avoid feeding the model large blocks of text or long paragraphs, as it is fine-tuned to prioritize conversational "taste" over document-style translation.

[!TIP] Parameter Optimization: To achieve the best results for your specific use case, we recommend fine-tuning your inference settings:

Temperature: Use 0.3 for high accuracy or 0.6 - 0.8 for more natural/Egyptian flair.

Repeat Penalty: Setting this around 1.1 helps keep the dialogue snappy and avoids formal repetition.

🆚 Style Comparison

Source English	Standard Translation (Literal)	Elgizawy Cinematic Style
"sam,I told you not to come here."	"سام، أخبرتك أن لا تأتي هنا."	"ـ (سام)، قلت لك ألّا تأتي إلى هنا"

📦 Option 1: Usage with MLX (Apple Silicon)

Best for running natively on Mac M-series chips.

📦 Installation & Setup

1. For MLX (Native Apple Silicon)

Best for running natively on Mac M-series chips (M1, M2, M3, M4, M5) with high efficiency.

pip install mlx-lm

2. For GGUF (Universal / Python Integration)

To run the GGUF version via Python, you need the llama-cpp-python library. Choose the command based on your hardware:

Apple Silicon (Metal Optimized):

CMAKE_ARGS="-DGGML_METAL=on" pip install llama-cpp-python

Windows/Linux (NVIDIA GPU/CUDA):

bash

# PowerShell
$env:CMAKE_ARGS="-DGGML_CUDA=on"; pip install llama-cpp-python

Linux/Bash

CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python

Universal (CPU Only):

pip install llama-cpp-python

3. For GUI & CLI Tools

This model is compatible with various interfaces:

LM Studio: Drag and drop the .gguf file into the application to start chatting.
Ollama: Create a Modelfile pointing to the GGUF and use ollama create.
Llama.cpp: Run directly via terminal:
./main -m model.gguf

Abdullah-afify/Elgizawy-EN_AR-translation-style-qwen-3-8b-Q4_K_M

🎬 Qwen-3-8B Cinematic Translator (English to Arabic)