Mtrini-SVL-1.1
~8B open-weight model on Qwen3-VL-8B-Instruct, trained for reasoning, coding, mathematics and Moroccan Darija.
Mtrini-SVL-1.1 adapts Qwen3-VL-8B-Instruct with LoRA, trained on roughly 58.6K unique examples spanning code reasoning, math instruction data and a Moroccan Darija QA and reasoning set.
SVL stands for Self-Verifying Loop — Compiwer AI's development concept of encouraging a model to check and refine its answers (think, verify, improve, repeat). The card describes SVL as an ongoing research direction.
ATU-1.1 score
6.67%
Specifications
- Status
- Released
- Base model
- Qwen3-VL-8B-Instruct (opens in new tab)
- Parameters
- ~8B
- Architecture
- Qwen3-VL
- Method
- LoRA / PEFT
- Precision
- BF16
- Training context
- 4,096 tokens
- Languages
- en, ar, fr, ary
- Focus
- Reasoning, Coding, Mathematics, Moroccan Darija
Downloads
Training data
- Moroccan Darija QA + Reasoning
- 2,170
- OpenCodeReasoning-2
- 30,000
- OpenMathInstruct-2
- 30,000
- Unique examples after deduplication
- 58,574
Training run
- Steps
- 3,661 (1 epoch)
- Training loss
- 0.5855
- Mean token accuracy
- 83.55%
Usage
llama-cli \
-m Mtrini-SVL-1.1-BF16.gguf \
-p "Explain artificial intelligence in Moroccan Darija."Limitations
- Experimental: can hallucinate, write incorrect code, misunderstand Darija or produce wrong mathematical answers.
- Training metrics describe the training run and are not standardized benchmark results.