🐪 Mtrini-SVL-1.1

Small lab. Big ambitions.

Mtrini-SVL-1.1 is an 8B model adaptation developed by Compiwer AI, built on Qwen3-VL-8B-Instruct and trained with a focus on reasoning, coding, mathematics, and Moroccan Darija.

🧠 SVL — Self-Verifying Loop

SVL = Self-Verifying Loop

The idea:

Think → Verify → Improve → Repeat

SVL is Compiwer AI's approach to encouraging models to check and refine their responses instead of immediately stopping at the first answer.

📚 Training Data

The final dataset contained approximately 58.6K unique examples:

• 🇲🇦 Moroccan Darija QA + Reasoning — 2,170
• 💻 OpenCodeReasoning-2 — 30,000
• 🧮 OpenMathInstruct-2 — 30,000

⚙️ Model Details

Property	Value
Model	Mtrini-SVL-1.1
Base	Qwen3-VL-8B-Instruct
Parameters	~8B
Architecture	Qwen3-VL
Training	LoRA / PEFT
Precision	BF16
Training Context	4096
Developer	Compiwer AI
Origin	🇲🇦 Salé, Morocco

🧪 Training

• 3,661 training steps
• 1 epoch
• Training loss: 0.5855
• Mean token accuracy: 83.55%

These are training metrics and are not standardized benchmark results.

📦 Repository

This repository contains the LoRA adapter for Mtrini-SVL-1.1.

For the merged model:

CompiwerAI/Mtrini-SVL-1.1-Merged

For local GGUF inference:

CompiwerAI/Mtrini-SVL-1.1-GGUF

⚠️ Limitations

Mtrini-SVL-1.1 is an experimental model and can still hallucinate, make coding mistakes, misunderstand Darija, or produce incorrect mathematical answers.

Always verify important outputs.

🇲🇦 Built by Compiwer AI

Mtrini is part of Compiwer AI's effort to build open AI systems focused on:

Reasoning • Coding • Mathematics • Moroccan Darija • Open AI

Building AI For Everyone.