CompiwerAI/Mtrini-SVL-1.1-Merged

1---2language:3  - en4  - ar5  - fr6  - ary7license: apache-2.08library_name: transformers9base_model: Qwen/Qwen3-VL-8B-Instruct10pipeline_tag: text-generation11tags:12  - mtrini13  - compiwer-ai14  - qwen3-vl15  - reasoning16  - coding17  - mathematics18  - darija19  - moroccan-arabic20  - self-verifying-loop21---22 23# 🐪 Mtrini-SVL-1.124 25## The model is out of the lab.26 27Mtrini-SVL-1.1 is an open-weight ~8B model developed by **Compiwer AI**, based on **Qwen3-VL-8B-Instruct**.28 29It was trained with a focus on:30 31- 🧠 Reasoning32- 💻 Coding33- 🧮 Mathematics34- 🇲🇦 Moroccan Darija35- 🌍 Multilingual tasks36 37And yes...38 39**It's a Mtrini. 🐪**40 41## 🔄 Self-Verifying Loop42 43**SVL = Self-Verifying Loop**44 45Our development concept:46 47```text48THINK49  ↓50CHECK51  ↓52IMPROVE53  ↓54REPEAT 🔄55 56The goal is to encourage the model to verify and refine its responses.57 58SVL is an ongoing research direction at Compiwer AI.59 60📚 Training61 62The final dataset contained 58,574 unique examples.63 64Dataset	Examples65🇲🇦 Moroccan Darija QA + Reasoning	2,17066💻 OpenCodeReasoning-2	30,00067🧮 OpenMathInstruct-2	30,00068⚡ Model Specs69Property	Value70Model	Mtrini-SVL-1.171Base	Qwen3-VL-8B-Instruct72Parameters	~8B73Architecture	Qwen3-VL74Training	LoRA / PEFT75Precision	BF1676Training Context	4096 tokens77Dataset	58,574 unique examples78Developer	Compiwer AI79Origin	🇲🇦 Salé, Morocco80📊 Training Run81Steps: 3,66182Epochs: 183Training Loss: 0.585584Mean Token Accuracy: 83.55%85 86These metrics describe the training run and should not be treated as standardized benchmark scores.87 88🚀 Usage89from transformers import Qwen3VLForConditionalGeneration, AutoTokenizer90 91model_id = "CompiwerAI/Mtrini-SVL-1.1-Merged"92 93tokenizer = AutoTokenizer.from_pretrained(94    model_id,95    trust_remote_code=True96)97 98model = Qwen3VLForConditionalGeneration.from_pretrained(99    model_id,100    torch_dtype="auto",101    device_map="auto",102    trust_remote_code=True103)104⚠️ Limitations105 106Mtrini is an experimental open model.107 108It can hallucinate, generate incorrect code, misunderstand prompts, or produce incorrect mathematical answers.109 110Verify important outputs.111 112🌍 Why Mtrini?113 114AI shouldn't only be built by the biggest companies on Earth.115 116Mtrini is Compiwer AI's experiment in building open AI with a focus on:117 118🇲🇦 Moroccan language119🧠 Reasoning120💻 Coding121🧮 Mathematics122🤝 Open AI123 124Building AI For Everyone.125 126Built in Salé, Morocco 🇲🇦