Instructions to use vext-ma/silma-darija-9b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use vext-ma/silma-darija-9b-lora with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("silma-ai/SILMA-9B-Instruct-v1.0")
model = PeftModel.from_pretrained(base_model, "vext-ma/silma-darija-9b-lora")

Transformers

How to use vext-ma/silma-darija-9b-lora with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="vext-ma/silma-darija-9b-lora")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("vext-ma/silma-darija-9b-lora", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use vext-ma/silma-darija-9b-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "vext-ma/silma-darija-9b-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "vext-ma/silma-darija-9b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/vext-ma/silma-darija-9b-lora

SGLang

How to use vext-ma/silma-darija-9b-lora with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "vext-ma/silma-darija-9b-lora" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "vext-ma/silma-darija-9b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "vext-ma/silma-darija-9b-lora" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "vext-ma/silma-darija-9b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use vext-ma/silma-darija-9b-lora with Docker Model Runner:
```
docker model run hf.co/vext-ma/silma-darija-9b-lora
```

You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

🇲🇦 DerejLLM Sovereign 9B — Moroccan Darija LoRA (SILMA Base)

DerejLLM 9B is a Sovereign AI adapter for Moroccan Darija built on SILMA-9B-Instruct-v1.0, fine-tuned with QLoRA v3 Master on curated Moroccan civic, legal, monetary, and cultural data.

This adapter is best used as part of the derej.Client() Compound AI system — which combines this neural backbone with deterministic Moroccan engines for zero-hallucination performance.

🏆 Benchmark Results

Zero-Hallucination Red-Team Audit (29 adversarial scenarios)

Metric	Score
Hallucination Rate	0.0%
Pass Rate	100.0% (29/29)
Fake law gating	✅ 4/4
Mythical entity gating	✅ 2/2
Fictitious institution gating	✅ 2/2
Toxic safety gating	✅ 2/2
Moroccan monetary math	✅ 5/5
Civic/legal truth	✅ 7/7
OOD epistemic humility	✅ 1/1

Out-of-Distribution Benchmark (6 unseen Moroccan scenarios)

System	Score
Raw Neural Model (2B)	0/6 (0.0%)
DerejLLM Compound System	6/6 (100.0%)

Darija HellaSwag-25

Metric	Score
HellaSwag Accuracy	100.0% (25/25)

SFT Training — Bench-200

Metric	Score
Bench-200 Accuracy	75.0%
vs Qwen-2B baseline	+35 points

🔧 Quick Start

from derej import Client

client = Client()
response = client.chat("واش عندي الحق نجدد البيرمي ديالي من بنك البريد؟")
print(response)

Raw adapter (without Compound AI system)

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("silma-ai/SILMA-9B-Instruct-v1.0")
tokenizer = AutoTokenizer.from_pretrained("silma-ai/SILMA-9B-Instruct-v1.0")
model = PeftModel.from_pretrained(base, "vext-ma/silma-darija-9b-lora")
model = model.merge_and_unload()

🏗️ Architecture — Compound AI System

derej.Client.chat(prompt)
    │
    ├── Step 0: Sovereign Epistemic Shield
    │     ├── Fake law detection (19 verified laws only)
    │     ├── Mythical entity gating
    │     ├── Fictitious institution gating
    │     └── Hazardous practice interception
    │
    ├── Step 1: Tool Routing (ONCF, BAM, CNSS)
    ├── Step 2: Moroccan Monetary Engine (MAD/Ryal/Centimes)
    ├── Step 3: Domain Handlers (civic, legal, tech, culinary, automotive)
    └── Step 4: Graph-RAG + ColBERT Fallback
          ├── Knowledge Graph (18 legal entities, multi-hop BFS)
          └── Speculative Semantic Cache (cosine threshold 0.85)

📊 Training Details — Sovereign v3 Master

Parameter	Value
Base model	silma-ai/SILMA-9B-Instruct-v1.0
Method	QLoRA (4-bit NF4)
LoRA rank	32
LoRA alpha	64
Training data	Moroccan Darija Sovereign v3 Master + v4 Frontier
Dataset size	~45,000+ curated examples
Training steps	~2,900
Training platform	Kaggle GPU (T4 x2)

🔒 Privacy

No personal audio, voices, or biometric data in this repository
No private training optimizer states (checkpoint optimizer.pt / rng_state removed)
Community users bring their own voices for zero-shot TTS cloning
Only model weights (adapter_model.safetensors) and tokenizer config are published

📜 Citation

@misc{vext2026silma,
  title={DerejLLM: A Sovereign Compound AI System for Moroccan Darija},
  author={Vext Research Team},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/vext-ma/silma-darija-9b-lora}
}

Downloads last month: 4

Model tree for vext-ma/silma-darija-9b-lora

Base model

silma-ai/SILMA-9B-Instruct-v1.0

Adapter

(2)

this model

vext-ma
/

silma-darija-9b-lora