silma-darija-kashif-2b-gguf

SILMA-2B-Darija GGUF: Quantized Moroccan Language Model

4-bit quantized GGUF weights for Ollama, llama.cpp, and local inference

SILMA-2B-Darija GGUF is a 4-bit quantized build of silma-2b-darija, fine-tuned for Moroccan Arabic (Darija / ary) and Arabizi. It is configured for offline and local execution on standard consumer hardware (Apple Silicon Macs, Linux, and Windows) using Ollama, LM Studio, or llama.cpp.

---

⚡ Quickstart with Ollama (1-Command Run)

Run directly using Ollama from Hugging Face:

ollama run hf.co/vext-ma/silma-darija-kashif-2b-gguf

Or build and create locally from the included Modelfile:

ollama create silma-darija -f Modelfile
ollama run silma-darija

---

💻 Quickstart with LM Studio / Jan / llama.cpp

1. Download silma-darija-kashif-2b-q4_k_m.gguf (1.6 GB).
2. Drop it into your LM Studio models folder (~/.cache/lm-studio/models/) or Jan.
3. Select the Gemma 2 chat template:
   • System prompt: Prepend to user turn
   • User prefix: <start_of_turn>user\n
   • Model prefix: <start_of_turn>model\n
   • Stop token: <end_of_turn>

Using llama-cli directly:

llama-cli -m silma-darija-kashif-2b-q4_k_m.gguf \
  -p "<bos><start_of_turn>user\nسلام كيداير لاباس عليك؟<end_of_turn>\n<start_of_turn>model\n" \
  -n 256 --temp 0.7 --top-p 0.9 --repeat-penalty 1.15

---

📊 Quantization Details

---

🇲🇦 Core Capabilities

• Moroccan Darija Fluency: Conversational Darija with regional vocabulary.
• Arabizi Comprehension: Direct understanding of Latin-script Darija (3 for ع, 7 for ح, 9 for ق, 5 for خ).
• Cultural & Administrative Knowledge: Answers questions on Moroccan history, traditions, cooking, and administrative procedures.
• Affective Prosody Tags: Supports speech inflection tags ([calm], [excited], [whisper]) for neural TTS integration.

---

❓ Frequently Asked Questions (FAQ)

How do I run this model with Ollama?

Run ollama run hf.co/vext-ma/silma-darija-kashif-2b-gguf directly in your terminal. Ollama will fetch the GGUF file and launch an interactive chat prompt.

Can I run this offline without internet access?

Yes. Once downloaded, the model executes entirely offline on your local CPU or GPU without sending data externally.

What are the minimum system requirements?

A standard laptop with at least 4 GB of system RAM is sufficient. On Apple Silicon Macs (M1/M2/M3/M4), inference runs at 35+ tokens per second.

🙏 Acknowledgements

• AtlasIA DODa Team: For creating and open-sourcing the foundational Moroccan Darija speech corpus and lexical database.
• SWivid / F5-TTS Team: For the non-autoregressive Flow Matching DiT architecture.
• Habibi-TTS Authors: For dialectal Arabic foundation speech models.
• @Muno459: For community testing, audio verification, and feedback on Moroccan Darija speech outputs.
• SILMA AI & Google Gemma Team: For the open foundation model backbones.
• llama.cpp & Ollama Teams: For the high-performance local inference engines.

---

📜 Citation & Credits

• Base Model: SILMA-Kashif-2B-Instruct-v1.0
• Fine-Tuning & Quantization: Vext Research Team
• Model Adapter: SILMA-Darija-2B-LoRA
• Live Demo: SILMA Darija Web Studio