OncoAgent: Open-Source Privacy-Preserving AI for Oncology Decision Support
The OncoAgent Research Group has shared a technical preprint about OncoAgent, an open-source clinical decision support system designed for oncology. It operates solely on an AMD Instinct MI300X GPU. This innovative system integrates a dual-tier fine-tuned language model architecture, featuring 9B and 27B parameter models, along with a multi-agent LangGraph structure. It also includes a four-stage Corrective RAG pipeline that adheres to over 70 NCCN and ESMO guidelines and a three-layer reflexion safety validator that upholds a strict Zero-PHI policy. The training involved 266,854 real and synthetic oncology cases, optimized using the Unsloth framework. The complete system is entirely open source, allowing for on-premises deployment. Limitations include a reliance on 36% synthetic cases and the need for clinical validation against certified oncologists. All resources will be available on Hugging Face Spaces and GitHub.
Key facts
- OncoAgent is an open-source, privacy-preserving clinical decision support system for oncology.
- It uses a dual-tier fine-tuned LLM architecture with 9B and 27B parameter models.
- The system integrates a multi-agent LangGraph topology with eight specialized nodes.
- A four-stage Corrective RAG pipeline covers 70+ NCCN and ESMO guidelines.
- A three-layer reflexion safety validator enforces a Zero-PHI policy.
- Training corpus: 266,854 real and synthetic oncological cases.
- Fine-tuning via QLoRA on AMD Instinct MI300X hardware using Unsloth framework.
- Sequence packing enabled full-dataset fine-tuning in ~50 minutes (56× acceleration).
- Human-in-the-loop gate for Tier 2 cases and low-confidence outputs.
- All code and model weights will be released on Hugging Face Spaces and GitHub.
Entities
Institutions
- OncoAgent Research Group
- National Comprehensive Cancer Network (NCCN)
- European Society for Medical Oncology (ESMO)
- AMD
- Hugging Face
- GitHub
- LangChain
- Unsloth
- BitsAndBytes
- Featherless.ai