Dataset

PCBA SMT Dataset

Our released dataset is built from two complementary sources:

📘 Standards-derived VQA

Constructed from information extracted from industrial standards. This subset focuses on standards knowledge, defect-cause analysis, and decision-making.

🏭 Real-world VQA

Built from actual factory production-line data. This subset focuses on real manufacturing imagery and inspection-oriented reasoning under practical visual variation.

💻Released Public Split

- Train: 8,200 questions

- Test: 7,200 questions

- Total: 15,400 questions

VQA Taxonomy

Participants evaluate their models across four VQA families:

📘 Standard-based Knowledge QA

Questions derived from industrial standards documents, covering standard interpretation, defect cause analysis, and handling decisions.

🔍 Factuality

Perception-centric tasks include component type recognition, mount-side identification, defect existence detection, and defect type recognition.

🔢 Quantitative Reasoning

Numeric short-answer questions for component quantity counting and pin/lead counting.

📐 Attribute Reasoning

Recognition of main component shape.