齿问大模型 OralGPT

OralGPT

🤖 OralGPT-7B-Preview

⚡ Introduction

Hello! Welcome to the repository for OralGPT (齿问大模型)!

OralGPT is a specialized medical LLM for intraoral photograph analysis. Through a customized curriculum learning design and DAPO training, we utilized 1,585 annotated intraoral photographs to enhance the performance of Qwen2.5-VL-7B-Instruct in this specific domain, achieving a level comparable to models such as GPT-4o, GPT-5, and Gemini2.5-Pro.

⚠️ OralGPT-7B-Preview is an early-stage model for investigating training approaches. It is not equipped with professional diagnostic competence, and any clinical application is strictly discouraged.

👨‍⚕️ Model

Model Access

OralGPT-7B-Preview is available on Huggingface:

	Parameters	Link
OralGPT-7B-Preview	7B	ModelScope Link
OralGPT-9B	9B	Available soon

Note: OralGPT-7B-Preview is based on Qwen2.5-VL-7B-Instruct.

Model Inference

Inference with OralGPT

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

processor = AutoProcessor.from_pretrained("FreedomIntelligence/OralGPT-7B-Preview")
model = Qwen2_5_VLForConditionalGeneration.from_pretrained("FreedomIntelligence/OralGPT-7B-Preview", torch_dtype="auto", device_map="auto")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "/path/to/your/image.png",
            },
            {"type": "text", "text": "请帮我分析这张患者的口内拍摄图片。"},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

🧐 Evaluation

Out-of-domain Evaluation on 18 oral disease lables

In this study, we trained OralGPT-preview on web-crawled dental images, each carefully annotated by experienced dental professionals into 18 categories.

We conduct the out-of-domain evaluation on AlphaDent dataset's test set, which was annotated with the same 18 categories by doctors, ensuring label consistency to reliably evaluate the model’s generalization across different data sources.

	Macro F1	Micro F1	Macro AUC	Micro AUC	Jaccard Index
Gemini-2.5-Pro	42.61	42.86	56.62	55.85	47.68
Doubao-1.5-Vision-Pro	32.80	41.33	59.83	61.10	60.55
GPT-4o	41.16	45.03	59.67	60.55	60.75
GPT-5	45.58	46.89	60.44	62.28	62.25
Qwen2.5-VL-7B-Instruct	33.16	39.28	53.57	55.50	56.80
OralGPT-7B-Preview	43.75	47.98	63.98	63.67	64.13

🎯 To-Do

• Commit the preview version of OralGPT
• Upload the technical report
• Open-source the training code
• Propose the paper
• Release the professional version of OralGPT

📖 About Us

We are from:

• The Chinese University of Hong Kong, Shenzhen 香港中文大学（深圳）
• Shenzhen Stomatology Hospital (Pingshan) of Southern Medical University 南方医科大学深圳口腔医院（坪山）
• Freedom AI 深圳自由动脉科技有限公司