数据集获取

git clone https://github.com/starmpcc/C-Reason.git

curl -L -o repo.zip https://github.com/starmpcc/C-Reason/archive/refs/heads/main.zip
unzip repo.zip

源站 README 摘录（使用方式）

C-Reason

• Official Implementation of the paper:
  Enhancing LLMs’’ Clinical Reasoning with Real-World Data from a Nationwide Sepsis Registry [arxiv]

TL;DR We propose enhancing the clinical reasoning capabilities of large language models (LLMs) by leveraging real-world clinical records.
By training on a sepsis registry, we developed \mname, which showed improved reasoning performance not only on the sepsis registry itself but also across other datasets.

Reproduction

• The training data comes from the sepsis registry, which is private, along with additional private datasets from the hospitalized cohort and stroke registry.
• We are currently discussing the potential release of the model.
• Of all the data used, only the MIMIC-III dataset (see Section 3.2) is publicly available.
  <details>
  <summary> Setup </summary>
• We use devcontainer for reproducibility.
• Modify .devcontainer/.env to set environment variables.
• Build the devcontainer using vscode or using devcontainer CLI
  <details>
  <summary> Sepsis Registry </summary>
• NOTE: This dataset is private.
• Preprocessing
  • python scripts/sepsis_preprocessing.py metadata_path {METADATA_PATH} data_path {DATA_PATH} kor_to_eng_dict_path {KOR_TO_ENG_DICT_PATH} output_path {OUTPUT_PATH}
• Training
  • accelerate launch config_file config/deepspeed.json main.py +launcher=grpo +dataset@dataset.grpo=grpo model_path=microsoft/phi-4 debug=false datset.grpo.data_path={OUTPUT_PATH}
• Evaluation
  • python main.py +launcher=evaluate debug=false +dataset@dataset.grpo=grpo dataset.grpo.data_path={OUTPUT_PATH} model_path={MODEL_PATH} dataset.sepsis_registry.all_victims=true dataset.sepsis_registry.num_eval_samples=99999999
    </details>

<details>
<summary> MIMIC-III </summary>

• Setup
  • git submodule init && git submodule update
  • cd ehr_postgres && python mimic_iii {MIMIC_III_PATH}
  • cd ../mimic_sepsis && python preprocess.py && python sepsis_cohort.py process_raw && cd ..
• Measurement Prediction Preprocessing
  • python scripts/mimic_all_victims.py data_path mimic_sepsis/sepsis_final_data_RAW_withTimes.csv output_path {OUTPUT_PATH}
• Mortality Prediction Preprocessing

数据加载示例（表格/文本类）

import pandas as pd, glob, os

files = (glob.glob(os.path.join(path, "**", "*.csv"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tsv"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.xlsx"), recursive=True))
print("数据文件:", files)
df = pd.read_csv(files[0])
print(df.shape); print(df.columns.tolist()); print(df.head(3))

完整仓库：github.com/starmpcc/C-Reason