lablab-ai-amd-developer-hackathon/OncoAgent-v1.0-27B

🧬 OncoAgent v1.0 — 27B (Tier 2)

Advanced Reasoning Model for Complex Oncology Cases

AMD Developer Hackathon 2026 · Deployed on AMD Instinct™ MI300X · ROCm 7.2

Model Description

OncoAgent v1.0 27B is the Tier 2 (advanced reasoning) model in the OncoAgent multi-agent oncology triage system. It leverages the full capacity of Qwen/Qwen3.6-27B with a specialized clinical oncology system prompt and RAG-grounded inference.

This model is activated for complex cases that require deeper reasoning:

• Multi-line therapy planning (Stage III/IV cancers)
• Rare tumor types with limited guideline coverage
• Cases requiring cross-guideline synthesis (NCCN + ESMO)
• Differential diagnosis with conflicting biomarkers

Architecture Role

In the OncoAgent dual-tier architecture, the 27B model is the "deep thinker":

Clinical Case → Router Agent
                    │
                    ├── Simple/Common → [Tier 1: 9B LoRA] → Fast Triage
                    │
                    └── Complex/Rare  → [Tier 2: 27B]     → Deep Analysis
                                              │
                                              ↓
                                        Specialist Agent
                                              │
                                              ↓
                                        Critic (Reflexion Loop)
                                              │
                                              ↓
                                     Validated Recommendation

Routing Criteria (Tier 1 → Tier 2 Escalation)

Trigger	Example
Stage III/IV disease	Metastatic breast cancer
Rare tumor types	Merkel cell carcinoma
Multi-drug regimens	Combination immunotherapy
Conflicting data	HER2-low with BRCA mutation
Low RAG confidence	Cross-encoder score < 0.70

Configuration

This model uses the base Qwen3.6-27B with OncoAgent's specialized system prompt and Corrective RAG pipeline. The configuration includes:

Parameter	Value
Base Model	Qwen/Qwen3.6-27B
Precision	BF16 (native MI300X Matrix Cores)
Context Window	32,768 tokens
Serving Engine	vLLM with PagedAttention
GPU Memory	~55% of MI300X 192GB HBM3
Tensor Parallelism	1 (single MI300X)

System Prompt

You are OncoAgent-Specialist, a board-certified oncologist AI assistant.
You provide evidence-based treatment recommendations grounded EXCLUSIVELY
in the retrieved clinical guidelines (NCCN/ESMO).

RULES:
1. NEVER invent treatments. If the evidence is not in the provided context,
   state: "Información no concluyente en las guías provistas."
2. Always cite the guideline source (NCCN/ESMO) and evidence category.
3. Structure your response with: Clinical Summary, Diagnostic Findings,
   Treatment Recommendation, and Evidence Level.
4. Consider comorbidities, contraindications, and patient-specific factors.
5. For Stage IV cases, include discussion of clinical trial eligibility.

vLLM Deployment (AMD MI300X)

# Serve Tier 2 on MI300X
python -m vllm.entrypoints.openai.api_server \
    --model Qwen/Qwen3.6-27B \
    --dtype bfloat16 \
    --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.55 \
    --max-model-len 32768 \
    --port 8001

Dual-Model Deployment

# Run both tiers simultaneously on MI300X (192GB HBM3)
# Tier 1 (9B): ~45% GPU memory → Port 8000
# Tier 2 (27B): ~55% GPU memory → Port 8001
bash deploy/start_vllm.sh both

Safety Features

OncoAgent v1.0 27B operates within a multi-layered safety framework:

1. Anti-Hallucination Policy — Model is constrained to RAG-retrieved context only
2. Reflexion Critic Loop — Output is validated by a dedicated Critic agent
3. Diagnostic Rigor Check — Treatment recommendations require confirmed pathology
4. PHI Sanitization — Zero patient health information in logs
5. HITL Gate — Stage IV cases can trigger human-in-the-loop review

Links

• 🔗 Demo: HF Space
• 🔗 GitHub: maximolopezchenlo-lab/OncoAgent
• 🔗 Tier 1 Model: OncoAgent-v1.0-9B
• 🔗 Dataset: OncoAgent-Clinical-266K

Citation

@misc{oncoagent2026,
  title={OncoAgent: Multi-Agent Oncology Triage System},
  author={Lopez Chenlo, Maximo},
  year={2026},
  howpublished={AMD Developer Hackathon 2026},
  url={https://github.com/maximolopezchenlo-lab/OncoAgent}
}

License

Apache 2.0 — This model configuration is for research and educational purposes only. Not intended for direct clinical use without professional medical oversight.