Qwen2.5-7B-Instruct-darija-gguf

Qwen2.5-7B-Instruct-darija-gguf

This repository contains quantized versions of Qwen2.5-7B-Instruct-darija in GGUF format for efficient inference.

Available Quantizations

Usage

Using llama.cpp

# Download the desired quantization
wget https://huggingface.co/GemMaroc/Qwen2.5-7B-Instruct-darija-gguf/resolve/main/Qwen2.5-7B-Instruct-darija_ckpt-*_q8_0.gguf

# Run inference
./llama-cli -m Qwen2.5-7B-Instruct-darija_ckpt-*_q8_0.gguf -p "Your prompt here"

Using Python with llama-cpp-python

from llama_cpp import Llama

# Load the quantized model
llm = Llama(
    model_path="./Qwen2.5-7B-Instruct-darija_ckpt-*_q8_0.gguf",
    n_ctx=32768,  # Context length
    n_threads=8,  # Number of CPU threads
)

# Generate text
response = llm("Your prompt here", max_tokens=512)
print(response['choices'][0]['text'])

Model Information

• Base Model: GemMaroc/Qwen2.5-7B-Instruct-darija
• Quantization: Multiple GGUF formats available
• Context Length: 32,768 tokens
• Languages: Arabic (Moroccan Darija), English

Recommendations

• For best quality: Use f16 (largest file size)
• For balanced performance: Use q8_0 (recommended)
• For resource-constrained environments: Use tq2_0 or tq1_0

Citation

If you use this model, please cite the original GemMaroc paper:

@misc{skiredj2025gemmarocunlockingdarijaproficiency,
      title={GemMaroc: Unlocking Darija Proficiency in LLMs with Minimal Data},
      author={Abderrahman Skiredj and Ferdaous Azhari and Houdaifa Atou and Nouamane Tazi and Ismail Berrada},
      year={2025},
      eprint={2505.17082},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.17082},
}