ULVR_v2: Universal Latent Visual Reasoning Dataset with ~1.51M Samples

~1.51 million samples across eight splits comprise this multimodal reasoning dataset. It was created by RuoliuYang and last updated on May 25, 2026. The splits include text-based chain-of-thought reasoning, bounding box manipulations, and visual representations like depth maps.

Use Cases

• Training visual question answering models based on text-to-image reasoning chains.
• Fine-tuning object detection models using bounding box crop and highlight augmentations.
• Developing multimodal reasoning agents based on interleaved image-text sequences.
• Benchmarking model performance on depth map and edge detection tasks.

Strengths

• Large scale with approximately 1.51 million total samples.
• Structured into eight distinct splits for different reasoning tasks, such as 'text_cot' with 355,793 samples.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count per split beyond the provided totals is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to the aggregated datasets like ScienceQA and LLaVA-CoT.

Provenance

Source

Aggregated from multiple datasets including ScienceQA, LLaVA-CoT, Mulberry-SFT, Visual-CoT, ThinkMorph, and others.

Collection Method

Likely compiled and processed from existing open-source datasets for visual reasoning.

Freshness

Last updated 2026-05-25 10:47:32; freshness should be verified.