Cancer-LLM: A PubMed Central–Derived Instruction Dataset and Fine-Tuned BioBERT Model for Multi-Class Cancer Type Classification

Cancer research produces a large volume of unstructured text, yet most general-purpose language models lack the focused, labeled data needed to reason about specific cancer types. We present Cancer-LLM, a biomedical text dataset and an accompanying classification model built around nine cancer types: blood, brain, breast, liver, lung, mouth/oral, ovarian, prostate, and skin. Source paragraphs were collected from open-access full-text articles in PubMed Central using the NCBI Entrez interface, then organized into three clinical categories: drugs and treatments, symptoms, and prevention. After cleaning, the material was reformatted into approximately 151,826 instruction-style examples suitable for language-model fine-tuning. Starting from BioBERT, a biomedical variant of BERT, we fine-tuned a model for multi-class cancer type classification and reached 89.77% accuracy across eight classes. The dataset and model are intended to support clinical AI assistants, retrieval-augmented oncology systems, and downstream biomedical NLP research. We describe the collection pipeline, the resulting dataset statistics, and the limitations of a keyword-driven extraction approach.