---
model_name: QwenMedic-v1
language: en
license: apache-2.0
pipeline_tag: text-generation
tags:
  - medical
  - clinical
  - question-answering
  - summarization
  - decision-support
datasets:
  - FreedomIntelligence/medical-o1-reasoning-SFT
  - jtatman/medical-sci-instruct-1m-sharegpt
---



## Model Card: QwenMedic-v1

<p align="center">
  <img
    src="https://huggingface.co/ross-dev/QwenMedic-v1/resolve/main/assets/model_image.png"
    alt="Model Image"
    width="350"
  />
</p>

### Overview  
QwenMedic-v1 is a medical-specialty adaptation of the Qwen3-1.7B causal language model, fine-tuned for clinical reasoning and instruction-following tasks. It was trained for **1 epoch** on two curated medical datasets to improve diagnostic Q&A and clinical summarization.

### Base Model  
- **Architecture:** Qwen3-1.7B (28 layers, 16 Q / 8 KV attention heads, 32 768-token context)  
- **Parameters:** 1.7 billion  
- **Quantization:** float16 and int4 supported  

### Fine-Tuning Data  
1. **Medical Reasoning SFT** (`FreedomIntelligence/medical-o1-reasoning-SFT`)  
   - Chain-of-thought reasoning examples on verifiable medical problems  
   - Language: English  
   - Split used: `train`  

2. **General Medical Instruction** (`jtatman/medical-sci-instruct-1m-sharegpt`)  
   - Conversational Q&A prompts across medical topics  
   - Sampled first 100 000 examples via `train[:100000]`  

### Training Configuration  
- **Framework:** PyTorch + Hugging Face Transformers  
- **Optimizer:** AdamW  
- **Learning Rate:** 2 × 10⁻⁵  
- **Batch Size:** 16 (with gradient accumulation)  
- **Precision:** bfloat16 mixed precision  
- **Hardware:** NVIDIA RTX 3090 (24 GB)  

### Intended Uses  
- Clinical question answering & differential diagnosis  
- Summarization of patient notes  
- Medical education & decision support  

### Limitations & Risks  
- May produce **hallucinations** or plausible-sounding but incorrect advice  
- **Biases** due to training-data coverage  
- **Not FDA-approved**—should not replace professional medical judgment  
- Avoid feeding **patient-identifiable** data without proper de-identification

### Summary of Final Training Metrics

| Metric            | Step | Smoothed | Raw Value |
|------------------:|-----:|---------:|----------:|
| **Epoch**         | 1539 | 0.9979   | 0.9997    |
| **Gradient Norm** | 1539 | 0.3882   | 0.3974    |
| **Learning Rate** | 1539 | —        | 0         |
| **Training Loss** | 1539 | 1.5216   | 1.4703    |

### Inference Example  

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/QwenMedic-v1"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "A 55-year-old male with Type 2 diabetes presents with sudden chest pain "
    "and diaphoresis. What are the top differential diagnoses?"
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)
```

### Contact
- **Creator:** Andre Ross
- **Company:** Ross Technologies
- **Email:** devops.ross@gmail.com