This paper deals with adjectival vagueness, introduced by a gradable adjective. In a legal domain presence of vague expressions is undesirable or explicitly forbidden. Our goal is to create a classifier for vagueness detection in Russian legal texts. A Russian dataset of vague expressions (words or sentences) did not exist, so we created one. We describe (1) previous experience in forming annotated datasets; (2) our experience in collecting and manually annotating sentences based on the presence/absence of vagueness criterion; and (3) the structure of the resulting dataset. Initially, we used multiple independent labeling by three lawyers. However, the legal team demonstrated a significantly different approaches to vagueness annotation (with inter-annotator agreement being less than the expected by chance), so we selected the most competent expert. Our resulting dataset is of 6,000 sentences. It contains contexts from the “CorCodex” corpus of laws. Each sentence is accompanied by lemmas with vagueness tags.