Synthetic Visual Reasoning Problems With 1M+ Samples

LGT is a synthetic dataset containing over 1 million visual reasoning problems paired with reasoning traces. It was created by NVIDIA for research and development, supporting the full VLM post-training spectrum including SFT and reinforcement learning.

Use Cases

• Fine-tune vision-language models using the 1M+ visual reasoning problems for supervised fine-tuning (SFT).
• Apply offline reinforcement learning (DPO) using the provided reasoning traces exhibiting complex cognitive behaviors.
• Train models for online reinforcement learning (GRPO/RLVR) with the dataset's spectrum of problem complexity.
• Analyze the reasoning chains to model and improve complex cognitive behaviors in AI systems.

Strengths

• Contains over 1 million new visual reasoning problems, providing significant scale for model training.
• Problems exhibit varying complexity, supporting a range of research and development tasks.
• Includes reasoning traces that model complex cognitive behaviors, adding depth to the data.
• Explicitly supports the entire VLM post-training spectrum (SFT, DPO, GRPO/RLVR).

Limitations

• The dataset is synthetic, which may limit its applicability to real-world visual reasoning scenarios.
• Intended for research and development only, restricting commercial applications.
• Specific column structure, file formats, and sample data are not provided in the input, limiting initial assessment.

Provenance

Source

NVIDIA

Collection Method

Synthetic generation of visual problems and reasoning chains.

Freshness

Last updated on March 17, 2026.