lablab-ai-amd-developer-hackathon/OncoAgent-Clinical-266K

OncoAgent临床数据集是一个包含266,854个临床肿瘤学训练样本的精选多源数据集，专门用于微调大型语言模型，以支持癌症诊断、治疗推荐和临床推理任务。该数据集汇集了来自PubMed Central的真实临床病例报告（PMC-Patients）、基于证据的医学问答（PubMedQA）、合成的链式思维肿瘤学推理对（OncoCoT）、85多种癌症类型的结构化治疗协议（NCCN指南提取物）以及欧洲临床实践指南（ESMO指南）。数据经过过滤、去重和质量控制，覆盖超过85种癌症类型，包括乳腺癌、肺癌、结直肠癌、前列腺癌等常见和罕见亚型。数据集采用兼容trl.SFTTrainer的聊天格式（JSONL），并分为训练集（240,168个样本）和评估集（26,686个样本）。所有数据均来自公开来源或合成生成，不含真实个人健康信息，仅供研究和教育用途。This dataset contains 266,854 clinical oncology training samples curated for fine-tuning large language models on cancer diagnosis, treatment recommendation, and clinical reasoning tasks. It aggregates real clinical case presentations from PubMed Central (PMC-Patients), evidence-based medical question answering (PubMedQA), synthetic chain-of-thought oncology reasoning pairs (OncoCoT), structured treatment protocols from over 85 cancer types (NCCN Guideline Extracts), and European clinical practice guidelines (ESMO Guidelines). The data is filtered, deduplicated, and quality-controlled, covering more than 85 cancer types including breast, lung, colorectal, prostate, and other common and rare subtypes. The dataset is formatted in chat-compatible JSONL for use with trl.SFTTrainer, with splits into training (240,168 samples) and evaluation (26,686 samples). All data is sourced from public materials or synthetically generated, contains no real protected health information, and is intended for research and educational purposes only.