README.md

---
base_model: unsloth/qwen3-vl-8b-instruct-unsloth-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- qwen3_vl
license: apache-2.0
language:
- en
- tr
datasets:
- ituperceptron/turkish-math-vqa
pipeline_tag: image-text-to-text
library_name: transformers
---

## Model Description

- Base Architecture: Qwen3-VL-8B-Instruct
- Fine-Tuning Method: QLoRA (PEFT)
- Language: Turkish
- Domain: High School Mathematics (12th Grade)
- Modality: Vision-Language (Image + Text → Text)

This model is a QLoRA fine-tuned version of Qwen3-VL-8B-Instruct trained on the Turkish-Math-VQA dataset, which consists of 12th-grade mathematics problems published by the Turkish Ministry of National Education (MEB).
The model is designed to:
- Understand mathematical problem images
- Generate step-by-step solutions in Turkish
- Handle topics such as logarithms, sequences & series, trigonometry, derivatives, and integrals


## Intended Use

Primary Use Cases
- Turkish mathematical Visual Question Answering (VQA)
- Educational AI assistants
- Step-by-step solution generation
- Math tutoring systems
- Research in Turkish multimodal reasoning


## Out-of-Scope Use
- Professional exam grading without human validation
- Safety-critical mathematical applications
- Guaranteed mathematically verified reasoning


## Training Data

**Dataset**: Turkish-Math-VQA
The dataset contains mathematics problems from official 12th-grade exams prepared by the Turkish Ministry of National Education.

**Dataset Fields**:
- `test_number`: The test identifier
- `question_number`: Question number within the test
- `image`: The image containing the math problem
- `solution`: Turkish solution generated synthetically using GPT-o1


**Important Note on Labels**:

The solution field was generated synthetically by GPT-o1 and has not been manually verified for correctness. While GPT-o1 is generally strong at solving problems at this level, the dataset may contain:
- Incorrect reasoning steps
- Logical inconsistencies
- Arithmetic mistakes

Therefore, the fine-tuned model may inherit these imperfections.

## How to Get Started with the Model

```python
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("khazarai/Math-VL-8B")
model = AutoModelForImageTextToText.from_pretrained("khazarai/Math-VL-8B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "Resimde verilen matematik problemini çözün."}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))
```

## Citation

If you use this model in academic work, please cite:
- The original Qwen model
- Turkish-Math-VQA dataset