Internal-dataset-variant-extraction is an English biomedical text dataset curated for structured genetic variant extraction. The dataset consists of manually selected passages from biomedical articles retrieved through LitVar, chosen to enrich the study corpus with additional examples of genetic variant mentions and their surrounding textual context. Each record contains a text passage and manually curated variant annotations represented as gene–variant pairs. The dataset is intended to support research on biomedical named entity recognition, genomic variant extraction, gene–variant association extraction, and structured literature mining with language models. In its current version, the dataset contains 20 passages, including 19 passages with annotated variants and 1 negative passage, with 180 annotated gene–variant pairs covering 159 unique variant strings and 19 unique gene symbols. The dataset was used as an internal resource in experiments on adapting small language models for structured genetic variant extraction.