|
--- |
|
library_name: transformers |
|
tags: [] |
|
--- |
|
|
|
# HumanF-MarkrAI/Gukbap-Qwen2.5-34B-VL🍚 |
|
|
|
## Model Details🍚 |
|
|
|
### Model Description |
|
- **Developed by:** HumanF-MarkrAI |
|
- **Model type:** Korean-VL-Qwen2.5-34B |
|
- **Language(s):** Korean + English |
|
- **Context Length:** 2048 |
|
- **License:** cc-by-nc-4.0 |
|
- **Finetuned from model:** [AIDC-AI/Ovis2-34B](https://huggingface.co/AIDC-AI/Ovis2-34B). |
|
|
|
### Model Sources |
|
When training, we used `H100 80GB GPU`x6. |
|
|
|
|
|
### Implications🍚 |
|
If you want to know our model's details, please see [🔥Gukbap-LMM Blog🔥](https://kyujinpy.tistory.com/169). |
|
And also, we provided the Korean-LMM training code based Ovis!! [🔥Github🔥](https://github.com/Marker-Inc-Korea/Ovis2-FFT-Korean). Please star⭐⭐!! |
|
|
|
|
|
### Training Method (SFT)🧐 |
|
The following papers contain the foundational methodologies for the dataset and training methods we are currently proceeding. |
|
- [LIMA](https://arxiv.org/abs/2305.11206). |
|
- [Ovis](https://arxiv.org/abs/2405.20797). |
|
|
|
|
|
### SFT Text-Datasets (Private) |
|
When we made the `Open-Source based dataset`, we use `microsoft/WizardLM-2-8x22B` through [DeepInfra](https://deepinfra.com/). |
|
Our datasets are made by `Evolving system`, which is propsed by [WizardLM](https://wizardlm.github.io/WizardLM2/). |
|
In training, we used 1849 training dataset, and 200 validation dataset. |
|
|
|
- **Wizard-Korea-Datasets:** [MarkrAI/Markr_WizardLM_train_ver4](https://huggingface.co/datasets/MarkrAI/Markr_WizardLM_train_ver4). |
|
> Learning rate: 2e-5; Epoch: 3 |
|
|
|
|
|
## Benchmakrs🤗 |
|
|
|
### Global MM Benchmark Score (Zero-shot) |
|
|
|
We internally evaluated [VLMEvalKit](https://github.com/open-compass/VLMEvalKit?tab=readme-ov-file). |
|
We utilized **chatgpt-0125**, **gpt-4o-mini** and **gpt-4-turbo** in `MMBench`, `MathVista` and `MMVet`, respectively. |
|
|
|
| Model | MMStar | MathVista | HallusionBench | AI2D | OCRBench | MMVet | MMBench_V11 | AVG | |
|
|:---------:|:-----:|:------:|:-----:|:-----:|:----:|:-----:|:-----:|:-----:| |
|
| Step-1o (closed model) | 69.3 | **74.7** | **89.1** | 55.8 | **92.6** | **82.8** | 87.3 | **78.8** | |
|
| InternVL2.5-78B-MPO (Open) | **72.1** | 76.6 | 58.1 | **89.2** | 90.9 | 73.5 | **87.8** | 78.3 | |
|
| Ovis2-34B (Open) | 69.2 | 76.1 | 58.8 | 88.3 | 89.4 | 77.1 | 86.5 | 77.9 | |
|
| InternVL2.5-38B-MPO (Open) | 70.1 | 73.6 | 59.7 | 87.9 | 89.4 | 72.6 | 85.4 | 77.0 | |
|
|:---------:|:-----:|:------:|:-----:|:-----:|:----:|:-----:|:-----:|:-----:| |
|
| **Gukbap-Qwen2-34B-VL🍚** | 69.33 | 77.40 | 55.66 | 88.31 | 84.7 | 74.13 | 86.53 | **76.58** | |
|
|:---------:|:-----:|:------:|:-----:|:-----:|:----:|:-----:|:-----:|:-----:| |
|
| Gemini-2.0-Flash | 69.4 | 70.4 | 58.0 | 83.1 | 82.5 | 73.6 | 71.0 | 72.6 | |
|
| GPT-4o-20241120 | 65.1 | 59.9 | 56.2 | 84.9 | 80.6 | 74.5 | 84.3 | 72.2 | |
|
| Ovis1.6-Gemma2-9B (Open) | 62.00 | 67.10 | 84.42 | 51.96 | 82.60 | 64.68 | 82.20 | 70.71 | |
|
| **Gukbap-Gemma2-9B-VL🍚** | 62.13 | 66.00 | 84.49 | 53.01 | 82.80 | 63.90 | 82.20 | **70.65** | |
|
| LLaVA-OneVision-72B | 65.8 | 68.4 | 47.9 | 86.2 | 74.1| 60.6 | 84.5 | 69.6 | |
|
| VARCO-VISION-14B (NCSoft) | 64.1 | 67.6 | 46.8 | 83.9 | 81.5 | 53.0 | 81.2 | 68.3 | |
|
| GPT-4o-mini-20240718 | 54.8 | 52.4 | 46.1 | 77.8 | 78.5 | 66.9 | 76.0 | 64.6 | |
|
> HallusionBench score: (aAcc + fAcc + qAcc) / 3 |
|
|
|
### Korean MM Benchmark Score (Zero-shot) |
|
|
|
We internally evaluated [🔥our code🔥](https://github.com/Marker-Inc-Korea/KoVLMEval). |
|
We utilized **gpt-4o-2024-08-06** in `K-LLAVA-W` evaluation. |
|
|
|
| Model | K-MMBench | K-MMStar | K-DTCBench | K-LLAVA-W | AVG | |
|
|:---------:|:-----:|:------:|:-----:|:-----:|:----:| |
|
| GPT-4o-20241120 | NaN | NaN | NaN | 85.50 | NaN | |
|
|:---------:|:-----:|:------:|:-----:|:-----:|:----:| |
|
| **Gukbap-Qwen2.5-34B-VL🍚** | 89.10 | 68.13 | 77.08 | **69.00** | **75.83** | |
|
| **Ovis2-34B** | **89.56** | **68.27** | 76.25 | 53.67 | 71.94 | |
|
| 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](https://rank.opencompass.org.cn/leaderboard-multimodal) |
|
- Korean MM Bench dataset: [NCSOFT](https://huggingface.co/NCSOFT). |
|
|
|
## Inference |
|
```python |
|
import torch |
|
from PIL import Image |
|
from transformers import AutoModelForCausalLM |
|
|
|
#import os |
|
#os.environ["cuda_visible_devices"]="0" |
|
|
|
# load model |
|
if __name__ == '__main__': |
|
# HumanF-MarkrAI/Gukbap-Qwen2-34B-VL |
|
# AIDC-AI/Ovis2-34B |
|
model = AutoModelForCausalLM.from_pretrained("HumanF-MarkrAI/Gukbap-Qwen2-34B-VL", |
|
torch_dtype=torch.bfloat16, |
|
multimodal_max_length=2048, |
|
cache_dir="/data/cache/", |
|
trust_remote_code=True).cuda() |
|
text_tokenizer = model.get_text_tokenizer() |
|
visual_tokenizer = model.get_visual_tokenizer() |
|
|
|
# single-image input (K-LLAVA-W) |
|
image_path = './images/ex_4.jpg' |
|
images = [Image.open(image_path)] |
|
max_partition = 9 |
|
text = '이미지에서 잘리지 않은 과일은 몇 개인가요?' |
|
query = f'<image>\n{text}' |
|
|
|
# format conversation |
|
prompt, input_ids, pixel_values = model.preprocess_inputs(query, images, max_partition=max_partition) |
|
attention_mask = torch.ne(input_ids, text_tokenizer.pad_token_id) |
|
input_ids = input_ids.unsqueeze(0).to(device=model.device) |
|
attention_mask = attention_mask.unsqueeze(0).to(device=model.device) |
|
if pixel_values is not None: |
|
pixel_values = pixel_values.to(dtype=visual_tokenizer.dtype, device=visual_tokenizer.device) |
|
pixel_values = [pixel_values] |
|
|
|
# generate output |
|
with torch.inference_mode(): |
|
gen_kwargs = dict( |
|
max_new_tokens=2048, |
|
do_sample=False, |
|
top_p=None, |
|
top_k=None, |
|
temperature=None, |
|
repetition_penalty=None, |
|
eos_token_id=model.generation_config.eos_token_id, |
|
pad_token_id=text_tokenizer.pad_token_id, |
|
use_cache=True |
|
) |
|
output_ids = model.generate(input_ids, pixel_values=pixel_values, attention_mask=attention_mask, **gen_kwargs)[0] |
|
output = text_tokenizer.decode(output_ids, skip_special_tokens=True) |
|
print(f'Output:\n{output}') |
|
``` |
|
|
|
## Chat Prompt😶🌫️ |
|
```yaml |
|
<|im_start|>user<image> |
|
Hello! My favorite food is Gukbap🍚!<|im_end|> |
|
<|im_start|>assistant |
|
(model answer) |
|
``` |
|
|
|
|
|
## Gukbap-VL Series models🍚🍚 |
|
- [HumanF-MarkrAI/Gukbap-Gemma2-9B-VL](https://huggingface.co/HumanF-MarkrAI/Gukbap-Gemma2-9B-VL) |
|
|
|
|
|
## BibTeX |
|
``` |
|
@article{HumanF-MarkrAI, |
|
title={Gukbap-Qwen2.5-34B-VL}, |
|
author={MarkrAI}, |
|
year={2025}, |
|
url={https://huggingface.co/HumanF-MarkrAI} |
|
} |
|
``` |
