MARVEL Component-Label OCR Dataset and Benchmarks (AIRISE Open Call 5)

Real-world smartphone photographs of electronic-component distributor labels (Mouser Electronics and DigiKey, Italian-locale prints) captured on reels and anti-static bags from two small-batch PCB production runs, with per-image part-number annotations verified against the distributors' own catalogue APIs (98.3%; 100% exact part-number match on all resolved labels). The dataset also includes two hand-adjudicated evaluation gold sets produced in the same experiment: both edge cases covering BOM parsing, component classification, value-code notation, designator ranges and functional-equivalence judgement, and grounded natural-language inventory questions (including must-refuse safety cases), together with the published benchmark result files comparing deterministic-heuristic and LLM-based engines. Produced by the MARVEL experiment (Buontech Solutions srl, Pisa, Italy) within AIRISE Open Call 5.