Add dataset card and metadata for ILVR/CoMT

README.md

@@ -0,0 +1,58 @@
+---
+license: mit
+task_categories:
+- image-text-to-text
+language:
+- en
+tags:
+- multimodal
+- reasoning
+- chain-of-thought
+---
+
+# CoMT (Chain of Multi-modal Thought) for ILVR
+
+This repository contains the processed CoMT (Chain of Multi-modal Thought) dataset used for the paper [Interleaved Latent Visual Reasoning with Selective Perceptual Modeling](https://huggingface.co/papers/2512.05665).
+
+[**Project Code**](https://github.com/XD111ds/ILVR) | [**Paper**](https://huggingface.co/papers/2512.05665)
+
+## Introduction
+Interleaved Latent Visual Reasoning (ILVR) is a framework that unifies dynamic state evolution with precise perceptual modeling. It interleaves textual generation with latent visual representations that act as specific, evolving cues for subsequent reasoning. The CoMT dataset is used to train and evaluate this paradigm, focusing on sequential multimodal reasoning.
+
+## Data Structure
+The dataset is organized into training and testing splits, along with associated image files:
+- `TRAIN.jsonl`: Training samples.
+- `TEST.jsonl`: Testing samples.
+- `comt.tar.gz`: Contains the image files required for the tasks.
+
+### Data Format
+The dataset follows the JSONL format with the following fields:
+- `text_input`: The question or instruction.
+- `image_input`: Initial input images.
+- `sequence_plan`: The interleaved chain-of-thought rationale containing "text" and "helper_image" paths.
+
+### Directory Setup
+To use this data with the [official ILVR implementation](https://github.com/XD111ds/ILVR), organize the directory as follows:
+```text
+data/
+├── TRAIN.jsonl
+├── TEST.jsonl
+└── images_comt/      <-- Extracted from comt.tar.gz
+    ├── creation/
+    └── ...
+```
+
+## Citation
+If you find this dataset or the ILVR framework useful, please cite the following paper:
+
+```bibtex
+@article{dong2025interleaved,  
+      title={Interleaved Latent Visual Reasoning with Selective Perceptual Modeling}, 
+      author={Shuai Dong and Siyuan Wang and Xingyu Liu and Zhongyu Wei},
+      year={2025},
+      eprint={2512.05665},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2512.05665}, 
+}
+```