Анализ результатов ML-экспериментов

★ 7.1 · data

analyze-results is a Claude Code skill that analyzes ML experiment results by computing statistics, building comparison tables, and generating structured insights. It scans figures/, results/, and other project output directories for JSON and CSV files, then organizes data by independent variables — model type, hyperparameters, data config — alongside dependent metrics such as perplexity, accuracy, and loss, always computing deltas against a baseline. With multiple seeds it reports mean ± std and checks reproducibility; when sweeping a parameter it identifies monotonic, U-shaped, or plateau trends and flags outliers. Each finding follows a four-part structure: observation with numbers, interpretation, implication for the research question, and a suggested next experiment. The skill suits ML engineers and researchers who need to quickly interpret a batch of runs and decide where to go next.