Model Description
This model was fine-tuned from Qwen2.5-VL-3B-Instruct using Parameter-Efficient Fine-Tuning (LoRA).
The training objective focused on:
- Mathematical reasoning
- Chain-of-thought style explanations
- Visual question answering
- ScienceQA mathematical problems
- Multimodal image-text understanding
- Turkish and English instruction following
Base Model
Fine-Tuning Method
- QLoRA (4-bit)
- LoRA adapters
- PEFT
- Hugging Face Transformers
Training Dataset
ScienceQA-Math-CoT
The dataset contains:
- ScienceQA mathematical questions
- Associated images
- Step-by-step solutions
- Final answers
The model was trained to generate reasoning traces before producing final answers.
Intended Uses
Suitable Uses
- Educational assistants
- Mathematical tutoring
- Visual mathematical reasoning
- STEM learning applications
- Homework support
- Science question answering
- Turkish and English multimodal assistants
Out-of-Scope Uses
This model is not intended for:
- Medical diagnosis
- Legal advice
- Financial decision-making
- Safety-critical systems
- Autonomous decision-making
Training Details
Hardware
- Kaggle Dual NVIDIA T4 GPUs
Training Configuration
Table with columns: Parameter, Value| Parameter | Value |
|---|
| Base Model | Qwen2.5-VL-3B-Instruct |
| Fine-Tuning | LoRA |
| Quantization | 4-bit NF4 |
| Precision | FP16 |
| LoRA Rank | 16 |
| LoRA Alpha | 32 |
| Learning Rate | 2e-4 |
| Batch Size | 1 |
| Gradient Accumulation | 4 |
Example Usage
from transformers import Qwen2_5_VLForConditionalGeneration
from transformers import AutoProcessor
from peft import PeftModel
base_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
"Qwen/Qwen2.5-VL-3B-Instruct",
device_map="auto"
)
model = PeftModel.from_pretrained(
base_model,
"salihfurkaan/Qwen2.5-VL-3B-ScienceQA-Math-CoT-Adapter"
)
processor = AutoProcessor.from_pretrained(
"Qwen/Qwen2.5-VL-3B-Instruct"
)
Limitations
- The model may still produce incorrect mathematical reasoning.
- Chain-of-thought outputs do not guarantee correctness.
- Performance depends on image quality and clarity.
- The model has not been evaluated on all mathematical domains.
- The model may hallucinate intermediate reasoning steps.
Ethical Considerations
- This model is intended for educational and research purposes.
- Users should independently verify mathematical solutions before relying on them in academic, professional, or real-world settings.
Citation
@misc{qwen2vl_scienceqa_math_cot,
title={Qwen2.5-VL-3B-ScienceQA-Math-CoT},
author={Salih Furkan Erik, Kerem Berke Başak},
year={2026},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/}}
}
