オープンソース腫瘍学エージェントOncoAgent、56倍の学習高速化を達成
プライバシー保護型の腫瘍学サポートエージェントOncoAgentは、AMD MI300Xハードウェア上で50分で学習され、API方式と比べて56倍の高速化を達成した。
the oncoagent research group published a technical preprint for oncoagent, a multi-agent framework designed for privacy-preserving oncology support. the system routes queries to 9b or 27b models fine-tuned via qdora on amd instinct mi300x hardware, processing 266,854 real and synthetic cases. hardware-specific optimizations enabled full dataset training in approximately 50 minutes, marking a 56× improvement over api-based generation. the architecture utilizes a four-stage corrective rag pipeline with 100% grading success and a reflexion safety validator to ensure data sovereignty.