Multimodal Scientific Reasoning & Visual QA Dataset | Bundled Hybrid Corpus

This dataset is a comprehensive, enterprise-grade multimodal reasoning resource designed to train and evaluate Large Multimodal Models (LMMs) on complex, multi-step scientific and technical tasks. Delivered as a unified hybrid bundle, it integrates real-world academic research and visual question-answering with a programmatically generated synthetic corpus. The synthetic component (10,000) is engineered to complement the core dataset by augmenting it with rare edge cases, controlled mathematical plot variations, and explicit triple-verification reasoning traces. Key Dataset Metrics Total Volume: 112,289 multi-step records across the bundled corpus Multimodal Content: Over 45,000 embedded Base64 Matplotlib diagrams, scientific figure references, and cross-document visual artifacts Structure: Delivered in JSONL format with complete reasoning chains, ground-truth answers, and executable generation code Validation: 100% schema and step-integrity pass rate across all records Core Capabilities & Features Advanced Graduate-Level Reasoning: Spans applied mathematics, differential equations, linear algebra, vector calculus, financial engineering, and quantum/physical systems. Programmatic Synthetic Augmentation: Features 13 distinct Matplotlib diagram types (phase portraits, contour maps, vector fields, and heatmaps) designed to expose models to challenging synthetic distribution shifts. Triple-Verification Architecture: Every synthetic sample includes a structured 5-step rea