LLaVA-CoT-100k: Vision-Language Reasoning Dataset

Introduced in the paper 'LLaVA-CoT: Let Vision Language Models Reason Step-by-Step', this dataset is designed to enable Vision-Language Models to perform autonomous multistage reasoning. It integrates 100,000 samples from various visual question-answering sources with structured reasoning annotations. The dataset was authored by Xkev and last updated on the Hugging Face platform in December 2025.

Use Cases

• Training VLMs for step-by-step reasoning based on the structured reasoning annotations.
• Benchmarking model performance on complex visual question-answering tasks.
• Studying methods to integrate multimodal data from diverse sources for autonomous reasoning.

Strengths

• Contains 100,000 samples, providing a substantial scale for model training.
• Integrates samples from various visual question-answering sources, suggesting diversity.
• Specifically designed with structured reasoning annotations to address systematic reasoning challenges.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Hugging Face dataset repository by author Xkev.

Collection Method

Integrated from various visual question-answering sources, with added reasoning annotations.

Freshness

Last updated 2025-12-20 08:27:29.