250,000 images sampled from the ImageNet-21K dataset form this multimodal reasoning collection. For each image, the dataset provides a prompt and two different step-by-step reasoning tokens and outputs. Created by krishnateja95 and last updated on October 8, 2025, it is designed for research on multimodal summarization.
Use Cases
- Training Vision Language Models on step-by-step reasoning based on the provided reasoning tokens.
- Evaluating model performance on multimodal summarization tasks using the paired prompts and answers.
- Benchmarking the ability of models to generate multiple valid reasoning paths for a single image.
Strengths
- Contains 250,000 images, providing a large-scale resource for training.
- Offers two distinct reasoning outputs per image, enabling evaluation of answer diversity.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
Provenance
- Source
- Images sampled from the ImageNet-21K dataset.
- Collection Method
- Synthetic generation of reasoning prompts and outputs for each image.
- Freshness
- Last updated 2025-10-08 21:39:20; freshness should be verified.