The first R1-styled model (ColonR1) tailored for reasoning in colonoscopy tasks

Figure 1: Details of our colonoscopy-specific reasoning model, ColonR1.

Quick start

• Before running the snippet, you need to install the following minimum dependencies.
  shell

  1conda create -n quickstart python=3.10 2conda activate quickstart 3pip install torch transformers accelerate pillow
• Then you can use python ColonR1/quickstart.py to run it, as shown in the following code.
  python

  1import torch 2from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration 3from PIL import Image 4import warnings 5import os 6 7warnings.filterwarnings('ignore') 8device = "cuda" if torch.cuda.is_available() else "cpu" 9 10MODEL_PATH = "ai4colonoscopy/ColonR1" 11IMAGE_PATH = "assets/example.jpg" 12Question = "Does the image contain a polyp? Answer me with Yes or No." 13 14print(f"[Info] Loading model from {MODEL_PATH}...") 15model = Qwen2_5_VLForConditionalGeneration.from_pretrained( 16 MODEL_PATH, 17 torch_dtype=torch.bfloat16, 18 attn_implementation="flash_attention_2", 19 device_map="auto" 20) 21model.eval() 22 23processor = AutoProcessor.from_pretrained(MODEL_PATH) 24 25if not os.path.exists(IMAGE_PATH): 26 raise FileNotFoundError(f"Image not found at {IMAGE_PATH}. Please provide a valid image path.") 27 28image = Image.open(IMAGE_PATH).convert("RGB") 29 30TASK_SUFFIX = ( 31 "Your task: 1. First, Think through the question step by step, enclose your reasoning process " 32 "in <think>...</think> tags. 2. Then provide the correct answer inside <answer>...</answer> tags. " 33 "3. No extra information or text outside of these tags." 34) 35 36final_question = f"{Question}

messages = [
	{
		"role": "user",
		"content": [
			{"type": "image", "image": IMAGE_PATH},
			{"type": "text", "text": final_question},
		],
	}
]

print("[Info] Processing inputs...")
text_prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = processor(
	text=[text_prompt],
	images=[image],
	padding=True,
	return_tensors="pt",
).to(device)


print("[Info] Generating response...")
with torch.no_grad():
	generated_ids = model.generate(
		**inputs,
		max_new_tokens=1024,
		do_sample=False
	)

generated_ids_trimmed = generated_ids[:, inputs.input_ids.shape[1]:]
output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True)[0]

print(output_text)
```

Reference

@article{ji2025colonx,
    title={Colon-X: Advancing Intelligent Colonoscopy from Multimodal Understanding to Clinical Reasoning},
    author={Ji, Ge-Peng and Liu, Jingyi and Fan, Deng-Ping and Barnes, Nick},
    journal={arXiv preprint arXiv:2512.03667},
    year={2025}
}

License