SciVision-5B

Abstract

SciVision-5B is a fine-tuned multimodal vision-language model designed for STEM reasoning and STEM diagram analysis. It is based on Qwen/Qwen3.5-4B and fine-tuned on a curated 20K-sample subset derived from MathV360K.

The training pipeline combines 4-bit NF4 QLoRA, dynamic input resolution, and a custom Multimodal Content Quality and Difficulty Scoring (MCQDS) curation algorithm. MCQDS prioritizes image-text pairs with explicit reasoning traces, mathematical notation, reasoning-oriented language, and structured question formats.

On a 240-sample evaluation subset spanning ScienceQA, MathVista, and MathVision, SciVision-5B achieves an average accuracy of 56.90%. Under the reported evaluation protocol.

The model is intended primarily for research in multimodal STEM reasoning, visual question answering, mathematical reasoning, and STEM diagram analysis.

Model Details

Property	Value
Model	SciVision-5B
Base Model	Qwen/Qwen3.5-4B
Parameters	~5B
Task	Image-Text-to-Text
Architecture	Multimodal Vision-Language Model
Fine-Tuning	4-bit NF4 QLoRA
Training Dataset	kerembb123/MathV360K-Candidate-20k
Source Dataset	Zhiqiang007/MathV360K
Input Resolution	Dynamic
Minimum Pixels	200,704
Maximum Pixels	301,056
License	Apache 2.0

Intended Use

SciVision-5B is intended for:

• Multimodal STEM reasoning
• STEM diagram analysis
• Mathematical visual question answering
• STEM figure and diagram interpretation
• Multiple-choice visual reasoning
• Image-grounded question answering
• Research on efficient multimodal VLM fine-tuning

The model is primarily intended for research and evaluation. Its outputs should not be assumed to be factually correct without independent verification.

Dataset Analysis & MCQDS Curation

Training Dataset

The primary training dataset is kerembb123/MathV360K-Candidate-20k, derived from Zhiqiang007/MathV360K.

A 20K candidate subset was selected and curated using the Multimodal Content Quality and Difficulty Scoring (MCQDS) algorithm.

The goal of MCQDS is to identify samples with higher supervision density for multimodal STEM reasoning instead of treating all image-text pairs as equally informative.

Curation Rationale

MCQDS was designed around the hypothesis that higher-quality multimodal supervision can be more valuable than indiscriminately increasing dataset size, particularly under constrained computational resources.

The curation strategy prioritizes examples containing:

1. Valid image-text alignment
2. Explicit reasoning traces
3. Mathematical notation
4. Reasoning-oriented language
5. Structured question or multiple-choice formats

Model Architecture & QLoRA

SciVision-5B is initialized from Qwen/Qwen3.5-4B.

The model retains the multimodal architecture of the base model and is adapted using parameter-efficient fine-tuning.

QLoRA Configuration

Fine-tuning uses 4-bit NormalFloat (NF4) quantization with QLoRA.

Configuration	Value
Quantization	4-bit NF4
LoRA Rank	16
LoRA Alpha	32
LoRA Dropout	0.05
Attention Targets	q_proj, k_proj, v_proj, o_proj
MLP Targets	gate_proj, up_proj, down_proj

The LoRA adapters target the principal attention projections and feed-forward linear modules.

Dynamic Input Resolution

Training uses dynamic image resolution with:

min_pixels = 200704
max_pixels = 301056

Dynamic resolution allows the model to retain more visual information when necessary while controlling memory consumption.

This is particularly relevant for STEM diagram analysis, where small labels, mathematical symbols, spatial relationships, and visual structures can affect the correct answer.

Training Dynamics

Hardware

Training was performed on:

2 × NVIDIA Tensor Core T4
16 GB VRAM per GPU

The combination of 4-bit quantization, QLoRA, and dynamic image resolution enabled fine-tuning within a constrained GPU memory budget.

Learning Rate

Training used a linear decay learning-rate schedule beginning at:

2.0e-4

The minimum reported training loss was 1.634, reached at approximately epoch 2.1.

Training Loss

Only the reported minimum loss and epoch are available. Intermediate values are not reconstructed or interpolated.

Training Loss
2.0 | ●
1.9 |  ●
1.8 |   ●
1.7 |    ●
1.6 |       ●  Minimum = 1.634
    +------------------------------
      0    0.5   1.0   1.5   2.1
                 Epoch

Training Metrics

Metric	Value
Initial Learning Rate	2.0e-4
LR Schedule	Linear Decay
Minimum Training Loss	1.634
Minimum-Loss Epoch	~2.1
Quantization	4-bit NF4
LoRA Rank	16
LoRA Alpha	32
LoRA Dropout	0.05
GPUs	2 × NVIDIA T4
GPU Memory	16 GB each
Minimum Image Pixels	200,704
Maximum Image Pixels	301,056

Hyperparameters

Hyperparameter	Configuration
Base Model	Qwen/Qwen3.5-4B
Fine-Tuning Method	QLoRA
Quantization	NF4, 4-bit
LoRA Rank	16
LoRA Alpha	32
LoRA Dropout	0.05
Learning Rate	2.0e-4
LR Scheduler	Linear Decay
Training Dataset	20K curated samples
Image Resolution	Dynamic
Min Pixels	200,704
Max Pixels	301,056
Hardware	2 × NVIDIA T4 16 GB

Benchmark Evaluation

Evaluation Protocol

SciVision-5B was evaluated on a 240-sample test subset covering:

• ScienceQA
• MathVista
• MathVision

The reported average is the arithmetic mean of the three benchmark accuracies.

Because the evaluation uses a 240-sample subset rather than the complete benchmark suites, these results should be interpreted as an experimental comparison rather than official full-benchmark scores.

Overall Results

Rank	Model	ScienceQA	MathVista	MathVision	Average
1	Claude 3.5 Sonnet	78.20%	73.90%	38.10%	63.40%
2	SciVision-5B (Ours)	71.80%	72.50%	26.40%	56.90%
3	Pixtral-12B	72.40%	58.00%	28.30%	52.90%
4	MiniCPM-V-2.6	70.10%	60.50%	25.10%	51.90%
5	Qwen2-VL-7B-Instruct	68.20%	58.20%	25.80%	50.73%
6	GPT-4V	71.00%	56.70%	21.40%	49.70%
7	LLaVA-1.5-13B	52.40%	37.80%	14.20%	34.80%
8	Random Baseline	25.00%	25.00%	20.00%	23.33%

SciVision-5B achieves 72.50% on MathVista, exceeding the reported MathVista scores of all other open-source models in the comparison set.

MathVista Detailed Metrics

Metric	Result
Evaluation Samples	240
Correct	174 / 240
Accuracy	72.50%
Weighted Precision	73.80%
Weighted Recall	72.50%
Weighted F1	73.12%

Performance Interpretation

The results indicate that targeted multimodal data curation combined with parameter-efficient adaptation can produce strong STEM reasoning performance with a relatively small model and training dataset.

SciVision-5B is particularly competitive on MathVista, where its 72.50% accuracy approaches the reported Claude 3.5 Sonnet result of 73.90%, while outperforming the evaluated open-source alternatives in this comparison.

These results should not be interpreted as definitive state-of-the-art claims across the complete benchmark suites because the evaluation is based on a 240-sample subset and comparison models may have been evaluated using different inference configurations.

Limitations

SciVision-5B has several important limitations.

First, the reported benchmark results are based on a 240-sample evaluation subset, not the complete official test sets. The resulting estimates may therefore have higher statistical variance than full-benchmark evaluations.

Second, MCQDS relies on heuristic signals such as word count, LaTeX density, keywords, and structured prompt markers. These signals are proxies for supervision quality and do not guarantee that every retained example contains correct or optimal reasoning.

Third, SciVision-5B inherits limitations from the underlying base model and its training distribution.

Fourth, the model can produce incorrect answers, hallucinated reasoning, or visually unsupported conclusions. Outputs should be independently verified in research, educational, or technical applications.

Finally, benchmark results are sensitive to prompt formatting, preprocessing, decoding parameters, and evaluation methodology. Direct comparisons should use identical evaluation protocols whenever possible.

Inference Code Snippet

The following example uses Hugging Face Transformers for multimodal inference.

import torch
from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "salihfurkaan/SciVision-5B"

processor = AutoProcessor.from_pretrained(MODEL_ID)

model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG",
            },
            {
                "type": "text",
                "text": (
                    "Analyze the STEM diagram carefully and answer the question. "
                    "Explain the reasoning before giving the final answer."
                ),
            },
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
    )

generated_tokens = outputs[0][inputs["input_ids"].shape[-1]:]

response = processor.decode(
    generated_tokens,
    skip_special_tokens=True,
)

print(response)

Note: The exact model-loading class and processor interface should match the Transformers implementation supported by the installed version and the base model configuration.

Example Prompting

For STEM diagram analysis:

Analyze the image carefully.

Identify the relevant visual information, equations, labels, and relationships.
Reason through the problem step by step.
Evaluate each answer choice when choices are provided.
Return the final answer clearly.

For multiple-choice tasks:

Analyze the STEM diagram and solve the problem.

Choices:
A. ...
B. ...
C. ...
D. ...

Explain your reasoning and finish with:

Final answer: <choice>

Reproducibility

The primary components required to reproduce the reported fine-tuning setup are:

Base model:
Qwen/Qwen3.5-4B

Dataset:
kerembb123/MathV360K-Candidate-20k

Fine-tuning:
4-bit NF4 QLoRA

LoRA:
r = 16
alpha = 32
dropout = 0.05

Learning rate:
2.0e-4

Scheduler:
Linear decay

Image resolution:
200,704–301,056 pixels

Hardware:
2 × NVIDIA T4 16 GB

The MCQDS scoring formulation is provided above to facilitate independent reproduction of the dataset curation procedure.

Citation

If you use SciVision-5B in academic research, please cite:

@misc{scivision5b2026,
  title        = {SciVision-5B: Efficient Multimodal STEM Reasoning with MCQDS-Curated QLoRA Fine-Tuning},
  author       = {Erik, Salih Furkan},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/salihfurkaan/SciVision-5B}},
  note         = {Fine-tuned vision-language model based on Qwen/Qwen3.5-4B}
}

Dataset Citation

@misc{mathv360k,
  title        = {MathV360K},
  author       = {Shen, Zhiqiang},
  year         = {2024},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ZhiqiangShen/MathV360K}}
}

License

SciVision-5B is released under the Apache License 2.0.

The model is derived from Qwen/Qwen3.5-4B. Users should review the applicable licensing and usage conditions of the base model and referenced datasets before redistribution or commercial deployment.

Acknowledgements

This work builds upon the Qwen3.5 multimodal architecture and the MathV360K dataset.

SciVision-5B was developed with an emphasis on efficient VLM adaptation under constrained computational resources, combining targeted multimodal data curation with parameter-efficient fine-tuning.

The model is intended to contribute to research on efficient multimodal STEM reasoning and STEM diagram analysis.