File size: 2,130 Bytes
272924b b36412c 272924b b36412c 272924b f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c f528977 b36412c 272924b b36412c 272924b b36412c 272924b b36412c 272924b b36412c 272924b b36412c 272924b b36412c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | ---
language:
- en
- ar
- fr
- ary
license: apache-2.0
library_name: transformers
base_model: Qwen/Qwen3-VL-8B-Instruct
tags:
- mtrini
- compiwer-ai
- qwen3-vl
- lora
- peft
- reasoning
- coding
- mathematics
- darija
- moroccan-arabic
---
# 🐪 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.** |