VGR-SFT: 10K-100K Visual Grounded Reasoning Samples for Multimodal SFT

VGR-SFT is a multimodal reasoning dataset of 10,000 to 100,000 records released by BytedanceDouyinContent in 2026. It integrates visual grounding with language-based deduction to facilitate explicit modeling of visual-region attention.

Use Cases

• Visual Question Answering using visual-region attention to ground text in image coordinates
• Supervised Fine-Tuning (SFT) of multimodal models to improve language-based deduction steps
• Evaluating autonomous attention mechanisms by comparing model outputs to grounded visual reasoning labels

Strengths

• 10,000 to 100,000 record scale
• CC0 1.0 public domain license
• Linked to ArXiv research paper 2506.11991

Limitations

• Sample data is unavailable for direct inspection
• Potential geographic bias as the dataset is hosted in the US region but authored by Bytedance
• Lack of explicit column documentation in the provided metadata

Provenance

Source

BytedanceDouyinContent, 'VGR: Visual Grounded Reasoning' (ArXiv:2506.11991)

Freshness

Last updated March 2026.

Geography

United States