BrainTumorClaims: A Dataset for Scientific Claim Detection and Structured Information Extraction in Brain Tumor Literature

BrainTumorClaims is a biomedical natural language processing (NLP) dataset for scientific claim mining and structured information extraction from literature on brain tumor research. The dataset contains sentence-level annotations extracted from scientific abstracts related to glioma, glioblastoma, neuro-oncology, medical imaging, radiomics, genomics, immunotherapy, radiotherapy, survival analysis, and associated biomedical research domains. The dataset supports two main tasks: 1. Claim Detection Sentence-level binary classification: * claim* non-claim 2. Structured Information Extraction For claim sentences, the following structured fields are annotated: * model* dataset* metric* value* task* domain* comparison The dataset follows strict annotation guidelines: * fields are annotated only when explicitly stated in the sentence.* implicit or inferred information is not added.* Unavailable information is represented as null. The project contains two complementary resources: * a large-scale training dataset containing LLM-assisted annotations with manual corrections,* and a manually validated golden benchmark dataset for validation and testing. Annotations were created using a hybrid workflow that combines: * LLM-assisted annotation,* manual correction,* expert review,* duplicate removal,* consistency validation. Several rounds of quality control were performed to improve annotation consistency, structured extraction reliability, sentence alignment, and JSONL validity. Dataset Characteristics * Sentence-level annotations* JSONL file format* DOI-linked scientific abstracts* Biomedical and neuro-oncology focused* Manually validated structured information* Compatible with transformer-based NLP pipelines Potential Applications * Scientific claim detection* Biomedical information extraction* Evidence mining* Automated literature review* Scientific knowledge graph construction* Neuro-oncology text mining* Benchmarking LLM-based extraction systems File Format The dataset is distributed in JSONL format. Each entry contains: * paper metadata,* sentence identifiers,* sentence text,* claim labels,* and structured scientific annotations, where applicable. Intended Usage The dataset is intended for: * training NLP models for claim extraction,* benchmark evaluation,* validation and testing,* biomedical information retrieval research,* An error analysis of scientific information extraction systems. Notes The dataset uses conservative annotation principles to reduce hallucinated annotations and improve reliability for biomedical NLP research and benchmarking.