Dataset for Syndrome Manifestation Recognition and Multidimensional Evaluation under Pathogenesis Reasoning Chain-of-Thought Supervision

This dataset was constructed for syndrome manifestation recognition in traditional Chinese medicine (TCM) under the framework of pathogenesis reasoning chain-of-thought (PR-CoT) supervision. It includes structured clinical cases, gold-standard syndrome labels, and predictions from multiple large language models (LLMs) across a primary spleen–stomach task and two zero-shot cross-disease settings (gynecology and heart-system disorders).The dataset does not involve explicit spatial attributes, and no fixed temporal resolution is defined. It consists of three CSV files. The predictions.csv file contains 700 samples, where each row represents a structured clinical case with model predictions organized in columns. The evaluations.csv file contains 5,400 records in a long-table format, where each row corresponds to a sample–model pair with associated multidimensional evaluation results. Evaluation metrics include structural parsing (Precision, Recall, F1 score, Jaccard similarity), semantic similarity (rated by independent LLMs), and expert review (consistency of syndrome differentiation and terminology standardization). The summary.csv file is used for generating figures and tables in the associated study.Missing values may occur when certain models fail to produce predictions; in such cases, prediction fields are left empty while available evaluation results are retained. Potential sources of error include variability in expert scoring, differences in LLM-based semantic evaluation, and limitations in structural parsing.This dataset provides a structured resource for evaluating LLM-based TCM syndrome manifestation recognition, supporting research on interpretable reasoning, semantic alignment, and cross-disease generalization.