HumanF-MarkrAI/Gukbap-Gemma3-12B-VL🍚

Model Details🍚

Model Description

• Developed by: HumanF-MarkrAI
• Model type: Korean-VL-Gemma3-12B
• Language(s): Korean + English
• Context Length: 2048
• License: cc-by-4.0
• Finetuned from model: google/gemma-3-12b-it.

Model Sources

When training, we used H100 80GB GPUx4.

Implications🍚

If you want to know our model's details, please see 🔥Gukbap-LMM Blog🔥.

Training Method (SFT)🧐

The following papers contain the foundational methodologies for the dataset and training methods we are currently proceeding.

• LIMA.
• Gemma3.

SFT Text-Datasets (Private)

When we made the Open-Source based dataset, we use microsoft/WizardLM-2-8x22B through DeepInfra.
Our datasets are made by Evolving system, which is propsed by WizardLM. In training, we used 1849 training dataset, and 200 validation dataset.

• Wizard-Korea-Datasets: MarkrAI/Markr_WizardLM_train_ver4.

  Learning rate: 1e-5; Epoch: 3

Benchmakrs🤗

Korean MM Benchmark Score (Zero-shot)

We internally evaluated 🔥our code🔥.
We utilized gpt-4o-2024-08-06 in K-LLAVA-W evaluation.

Model	K-MMBench	K-MMStar	K-DTCBench	K-LLAVA-W	AVG
Gukbap-Gemma3-12B🍚	82.88	48.53	64.17	72.83	67.10
gemma-3-12b-it	82.24	48.13	66.25	68.50	66.28
Gukbap-Gemma2-9B-VL🍚	80.16	54.20	52.92	63.83	62.78
Ovis1.6-Gemma2-9B	52.46	50.40	47.08	55.67	51.40
VARCO-VISION-14B	87.16	58.13	85.42	51.17	70.47
llama-3.2-Korean-Bllossom-AICA-5B	26.01	21.60	17.08	45.33	27.51

MM Benchmarks

• Global MM Bench dataset: OpenCampass MM leaderboard
• Korean MM Bench dataset: NCSOFT.

Gukbap-VL Series models🍚🍚

• HumanF-MarkrAI/Gukbap-Gemma2-9B-VL
• HumanF-MarkrAI/Gukbap-Gemma3-4B-VL
• HumanF-MarkrAI/Gukbap-Gemma3-12B-VL
• HumanF-MarkrAI/Gukbap-Gemma3-27B-VL

BibTeX

@article{HumanF-MarkrAI,
  title={Gukbap-Gemma3-12B-VL},
  author={MarkrAI},
  year={2025},
  url={https://huggingface.co/HumanF-MarkrAI}
}