Add dataset card for ComLesion-14K

README.md

@@ -0,0 +1,21 @@
+---
+task_categories:
+- image-segmentation
+tags:
+- medical
+- reasoning
+- lesion-segmentation
+---
+
+# ComLesion-14K
+
+ComLesion-14K is the first diverse Chain-of-Thought (CoT) benchmark for reasoning-driven complex lesion segmentation, introduced in the paper [CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning](https://huggingface.co/papers/2603.05911).
+
+- **Project Page:** [https://xyx1024.github.io/CORE-Seg.github.io/](https://xyx1024.github.io/CORE-Seg.github.io/)
+- **Repository:** [https://github.com/xyx1024/CORE-Seg](https://github.com/xyx1024/CORE-Seg)
+
+## Description
+
+Medical image segmentation is undergoing a paradigm shift from conventional visual pattern matching to cognitive reasoning analysis. ComLesion-14K addresses the gap in existing Multimodal Large Language Models (MLLMs) by providing specialized visual reasoning required for complex lesions. 
+
+The benchmark encourages models to integrate linguistic reasoning with pixel-level segmentation through a Chain-of-Thought approach. It was developed to evaluate the CORE-Seg framework, which utilizes a Semantic-Guided Prompt Adapter and reinforcement learning (GRPO) to improve both segmentation accuracy and logical interpretability in medical imaging.