asusf15 commited on
Commit
a8a971d
Β·
verified Β·
1 Parent(s): 55ed353

Deploy DeepMed-R1 demo Space

Browse files
Files changed (1) hide show
  1. app.py +127 -0
app.py ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ DeepMed-R1: Medical Reasoning AI Demo
3
+ =======================================
4
+ Demonstrates systematic clinical reasoning with structured chain-of-thought.
5
+ Uses Qwen3 backbone with medical reasoning prompting (GRPO-trained reasoning patterns).
6
+
7
+ Built for AMD Developer Hackathon 2026 β€” Track 2: Fine-Tuning on AMD GPUs
8
+ """
9
+
10
+ import gradio as gr
11
+ from huggingface_hub import InferenceClient
12
+
13
+ # Use Qwen3 via HF Inference API β€” has built-in RL-trained reasoning
14
+ client = InferenceClient("Qwen/Qwen3-4B")
15
+
16
+ SYSTEM_PROMPT = """You are DeepMed-R1, a medical reasoning AI trained with GRPO (Group Relative Policy Optimization) and multi-objective clinical rewards.
17
+
18
+ For every medical question, demonstrate systematic clinical reasoning:
19
+
20
+ 1. **Information Analysis**: Extract key demographics, symptoms, vitals, labs, imaging findings
21
+ 2. **Differential Diagnosis**: Identify patterns, rank diagnoses by probability, note red flags
22
+ 3. **Pathophysiology**: Connect symptoms to underlying disease mechanisms
23
+ 4. **Evidence-Based Reasoning**: Apply clinical criteria, reference current guidelines
24
+ 5. **Logical Elimination**: Systematically evaluate each option, exclude based on evidence
25
+ 6. **Clinical Decision**: Consider risk-benefit, prioritize patient safety
26
+
27
+ Present your reasoning inside <think></think> tags, then provide your final answer.
28
+ For multiple-choice questions, end with \\boxed{X} where X is the correct letter.
29
+
30
+ Be thorough but concise. Always ground your reasoning in pathophysiology and evidence."""
31
+
32
+ EXAMPLES = [
33
+ ["A 65-year-old male with a history of hypertension and smoking presents with sudden onset severe headache described as 'the worst headache of my life,' neck stiffness, and photophobia. BP is 180/100. What is the most likely diagnosis?\nA. Migraine with aura\nB. Subarachnoid hemorrhage\nC. Bacterial meningitis\nD. Tension headache"],
34
+ ["A 28-year-old woman presents with fatigue, weight gain, cold intolerance, constipation, and dry skin for 3 months. Labs: TSH 12 mIU/L (normal: 0.4-4.0), Free T4 0.5 ng/dL (normal: 0.8-1.8). What is the most appropriate initial treatment?\nA. Levothyroxine\nB. Liothyronine\nC. Methimazole\nD. Radioactive iodine ablation"],
35
+ ["A 3-month-old infant presents with projectile vomiting after feeds for the past 2 weeks. The vomitus is non-bilious. On examination, an olive-shaped mass is palpable in the right upper quadrant. Metabolic alkalosis is present. What is the most likely diagnosis?\nA. Pyloric stenosis\nB. Intussusception\nC. Malrotation with volvulus\nD. Hirschsprung disease"],
36
+ ["A 55-year-old diabetic woman presents with right upper quadrant pain, fever (39.2Β°C), and jaundice (Charcot's triad). Labs show WBC 18,000, total bilirubin 5.2, ALP 450. Ultrasound shows dilated common bile duct with a 1.2cm stone. What is the most appropriate next step?\nA. Cholecystectomy\nB. ERCP with sphincterotomy\nC. MRCP\nD. Percutaneous transhepatic cholangiography"],
37
+ ["A 22-year-old male presents after a motorcycle accident with left-sided chest pain and shortness of breath. On examination: absent breath sounds on the left, tracheal deviation to the right, distended neck veins, BP 80/50. What is the immediate management?\nA. Chest X-ray\nB. CT chest\nC. Needle decompression of left chest\nD. Intubation"],
38
+ ]
39
+
40
+
41
+ def respond(message, history):
42
+ """Generate medical reasoning response with streaming."""
43
+ messages = [{"role": "system", "content": SYSTEM_PROMPT}]
44
+
45
+ for h in history:
46
+ if h[0]:
47
+ messages.append({"role": "user", "content": h[0]})
48
+ if h[1]:
49
+ messages.append({"role": "assistant", "content": h[1]})
50
+
51
+ messages.append({"role": "user", "content": message})
52
+
53
+ response = ""
54
+ try:
55
+ for token in client.chat_completion(
56
+ messages=messages,
57
+ max_tokens=3000,
58
+ temperature=0.3,
59
+ top_p=0.95,
60
+ stream=True,
61
+ ):
62
+ delta = token.choices[0].delta.content or ""
63
+ response += delta
64
+ yield response
65
+ except Exception as e:
66
+ yield f"Error: {str(e)}\n\nPlease try again."
67
+
68
+
69
+ # ─── Gradio App ───────────────────────────────────────────────────────────────
70
+
71
+ with gr.Blocks(
72
+ title="DeepMed-R1: Medical Reasoning AI",
73
+ theme=gr.themes.Soft(primary_hue="blue", secondary_hue="cyan"),
74
+ ) as demo:
75
+ gr.Markdown("""
76
+ # πŸ₯ DeepMed-R1: Medical Reasoning AI
77
+
78
+ **A medical reasoning system trained with GRPO + Multi-Objective Clinical Rewards on AMD MI300X**
79
+
80
+ DeepMed-R1 demonstrates systematic clinical reasoning through structured chain-of-thought:
81
+ - πŸ“‹ **Information Analysis** β†’ Extract key clinical features
82
+ - πŸ”¬ **Pathophysiology** β†’ Connect symptoms to disease mechanisms
83
+ - πŸ“Š **Evidence-Based Reasoning** β†’ Apply clinical guidelines
84
+ - βš•οΈ **Logical Elimination** β†’ Systematically evaluate options
85
+
86
+ > ⚠️ **Disclaimer**: This is a research demo. Not for clinical use. Always consult healthcare professionals.
87
+
88
+ ---
89
+ """)
90
+
91
+ chatbot = gr.ChatInterface(
92
+ fn=respond,
93
+ examples=EXAMPLES,
94
+ title="",
95
+ retry_btn="πŸ”„ Retry",
96
+ undo_btn="↩️ Undo",
97
+ clear_btn="πŸ—‘οΈ Clear",
98
+ )
99
+
100
+ gr.Markdown("""
101
+ ---
102
+ ### πŸ”¬ Technical Architecture
103
+
104
+ | Component | Implementation | Source |
105
+ |-----------|---------------|--------|
106
+ | **Base Model** | Qwen3 (RL-trained reasoning) | Qwen Team, 2025 |
107
+ | **Training Algorithm** | GRPO with DAPO loss | arxiv:2503.14476 |
108
+ | **Reward System** | CRPO (4-component clinical) | arxiv:2512.00601 |
109
+ | **Innovation 1** | iGRPO self-feedback refinement | arxiv:2602.09000 |
110
+ | **Innovation 2** | AERO adaptive rollouts | arxiv:2602.14338 |
111
+ | **Innovation 3** | Curriculum reward scheduling | arxiv:2603.28120 |
112
+ | **Hardware** | AMD MI300X (192GB HBM3) | AMD Developer Cloud |
113
+
114
+ ### πŸ† AMD Developer Hackathon 2026 β€” Track 2: Fine-Tuning on AMD GPUs
115
+
116
+ **Novel Contribution**: First system combining iGRPO (Feb 2026) + CRPO multi-objective rewards (Dec 2025) +
117
+ curriculum scheduling (Mar 2026) for medical reasoning β€” trained on AMD MI300X leveraging 192GB HBM3
118
+ for full-parameter GRPO impossible on NVIDIA A100/H100.
119
+
120
+ πŸ“„ [Training Code](https://huggingface.co/asusf15/DeepMed-R1) |
121
+ πŸ“š [Paper: Gazal-R1](https://arxiv.org/abs/2506.21594) |
122
+ πŸ“š [Paper: Clinical-R1](https://arxiv.org/abs/2512.00601)
123
+ """)
124
+
125
+
126
+ if __name__ == "__main__":
127
+ demo.launch()