SpatialForge-10M: 10 Million Vision-Language QA Pairs for 3D Spatial Reasoning

SpatialForge-10M is a large-scale vision-language dataset designed for 3D-aware spatial perception and reasoning from open-world 2D images. It contains over 10 million question-answer pairs generated from 2.8 million curated images. The dataset was created by researchers from Lingnan University and XPENG Robotics, with the latest update recorded on 2026-05-25.

Use Cases

• Training vision-language models for 3D spatial reasoning based on 2D images.
• Benchmarking AI systems on spatial perception tasks using the generated QA pairs.
• Developing models for open-world scene understanding based on the curated image collection.

Strengths

• Contains over 10 million QA pairs, providing a large-scale resource.
• Built from a curated collection of 2.8 million 2D images.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the underlying image data is unknown, which may limit suitability assessment.
• Data may reflect geographic or source bias inherent to the open-world image collection.

Provenance

Source

Lingnan University · XPENG Robotics

Collection Method

Generated QA pairs from a curated collection of open-world 2D images.

Freshness

Last updated 2026-05-25 11:59:14; freshness should be verified.