Add model card and Apache-2.0 license
#3
by RobinChen2001 - opened
README.md
CHANGED
|
@@ -32,6 +32,7 @@ injection under a common grounded reasoning policy.
|
|
| 32 |
|
| 33 |
- **Base model:** [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)
|
| 34 |
- **Training data:** [IQuestLab/UniReason-Med-Data](https://huggingface.co/datasets/IQuestLab/UniReason-Med-Data)
|
|
|
|
| 35 |
- **Modalities:** image + text → text
|
| 36 |
- **License:** Apache-2.0
|
| 37 |
|
|
@@ -63,36 +64,6 @@ This checkpoint is the merged Hugging Face model exported from the GRPO stage.
|
|
| 63 |
Training code (LLaMA-Factory for SFT, verl for GRPO) and configs are released at:
|
| 64 |
<https://github.com/IQuestLab/unireason-med>.
|
| 65 |
|
| 66 |
-
## Usage
|
| 67 |
-
|
| 68 |
-
```python
|
| 69 |
-
from transformers import AutoModelForImageTextToText, AutoProcessor
|
| 70 |
-
from PIL import Image
|
| 71 |
-
|
| 72 |
-
model_id = "IQuestLab/UniReason-Med"
|
| 73 |
-
processor = AutoProcessor.from_pretrained(model_id)
|
| 74 |
-
model = AutoModelForImageTextToText.from_pretrained(model_id, device_map="auto")
|
| 75 |
-
|
| 76 |
-
image = Image.open("medical_image.png")
|
| 77 |
-
messages = [
|
| 78 |
-
{
|
| 79 |
-
"role": "user",
|
| 80 |
-
"content": [
|
| 81 |
-
{"type": "image"},
|
| 82 |
-
{"type": "text", "text": "What is the most likely diagnosis? Reason step by step."},
|
| 83 |
-
],
|
| 84 |
-
}
|
| 85 |
-
]
|
| 86 |
-
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 87 |
-
inputs = processor(text=[prompt], images=[image], return_tensors="pt").to(model.device)
|
| 88 |
-
output = model.generate(**inputs, max_new_tokens=1024)
|
| 89 |
-
print(processor.batch_decode(output, skip_special_tokens=True)[0])
|
| 90 |
-
```
|
| 91 |
-
|
| 92 |
-
The model produces interleaved reasoning with bounding boxes over the input image. Reproducing
|
| 93 |
-
the full grounded crop-and-continue loop (crop the predicted region and feed it back as visual
|
| 94 |
-
input) follows the agent/rollout logic in the released training code.
|
| 95 |
-
|
| 96 |
## Intended Use and Limitations
|
| 97 |
|
| 98 |
- **Intended use:** research on medical multimodal reasoning, visual grounding, and 2D-to-3D
|
|
|
|
| 32 |
|
| 33 |
- **Base model:** [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct)
|
| 34 |
- **Training data:** [IQuestLab/UniReason-Med-Data](https://huggingface.co/datasets/IQuestLab/UniReason-Med-Data)
|
| 35 |
+
- **Code:** [github.com/IQuestLab/unireason-med](https://github.com/IQuestLab/unireason-med)
|
| 36 |
- **Modalities:** image + text → text
|
| 37 |
- **License:** Apache-2.0
|
| 38 |
|
|
|
|
| 64 |
Training code (LLaMA-Factory for SFT, verl for GRPO) and configs are released at:
|
| 65 |
<https://github.com/IQuestLab/unireason-med>.
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
## Intended Use and Limitations
|
| 68 |
|
| 69 |
- **Intended use:** research on medical multimodal reasoning, visual grounding, and 2D-to-3D
|