c3med-reasoning

<div align="center"> # 🏥 C3Med-Reasoning **A Large-Scale Chinese Clinical Reasoning Dataset Re-labeled with Self-Reflection** *China Electronics Cloud Technology Co., Ltd.* <br/> <p align="center"> <img src="https://img.shields.io/badge/Size-1.5M%2B%20QA%20Pairs-27AE60" alt="Size"/> <img src="https://img.shields.io/badge/Storage-%3E10GB-E67E22" alt="Storage"/> <img src="https://img.shields.io/badge/Tasks-3%20Clinical%20Families-C0392B" alt="Tasks"/> <img src="https://img.shields.io/badge/Departments-17-8E44AD" alt="Departments"/> <img src="https://img.shields.io/badge/Annotation-3--Level%20Pipeline-4A90D9" alt="Annotation"/> <img src="https://img.shields.io/badge/License-Apache--2.0-16A085" alt="License"/> </p> </div> --- ## 📖 Overview **C3Med-Reasoning** is the training corpus behind the **C3Med** model — a large language model for Chinese-centered clinical medicine. It contains **more than 1.5 million question-answer pairs**, spanning **3 clinical task families** across **17 medical departments**, with every sample annotated with a **chain-of-thought (CoT) reasoning path**. Unlike conventional medical SFT corpora, where reasoning paths are borrowed from knowledge-graph inference or distilled from stronger teacher LLMs (which often contradict the target model's native thinking patterns), every CoT path in this dataset is **re-labeled by the target model itself** through an answer → verify → reflect loop. The resulting reasoning paths therefore embody sufficient domain knowledge while staying aligned with the model's own reasoning paradigm, making the corpus highly "tuning-friendly" and maximizing the effectiveness of subsequent supervised fine-tuning. The dataset was used to fine-tune Qwen3-8B into **C3Med-8B (v3)**, which achieves an average of **74.12% on the CMB benchmark** — **+42.6 points** over Baichuan-M2 (24B) — with only 8B parameters. --- ## 📦 Release Contents & Data Sources The dataset was assembled from **well-known open-source Chinese medical datasets** and re-labeled with the C3Med three-level annotation pipeline. All files are in JSON / JSONL format (UTF-8): | File | Source Dataset | Format | Approx. Samples | Size | | :--- | :------------- | :----: | :-------------: | ---: | | `Meds-Ins-250000_1.json` … `Meds-Ins-250000_4.json` | [MedS-Ins](https://github.com/WangRongming/MedS-Ins) | JSON | ~250,000 each | 4 × 1.6 GB | | `ReasonMed_352258_cleaned.json` | [ReasonMed](https://modelscope.cn/datasets/AiMed/ReasonMed) | JSON | ~352,000 | 856 MB | | `CMB_269303.json` | [CMB – Chinese Medical Benchmark](https://github.com/FreedomIntelligence/CMB) | JSON | ~269,000 | 612 MB | | `Huatuo26M_baichuan_176032.jsonl` | [Huatuo26M](https://github.com/FreedomIntelligence/Huatuo26M) (Baichuan-filtered subset) | JSONL | ~176,000 | 1.96 GB | | `UCSC_MedReason_32682.json` | UCSC MedReason | JSON | ~33,000 | 224 MB | | `CMExam_12654.json` | [CMExam](https://github.com/williamliwl/CMExam) | JSON | ~12,700 | 31 MB | | `MedQuad_16226.jsonl` | [MedQuad-MedicalQnADataset](https://github.com/abachaa/MedQuAD) | JSONL | ~16,200 | 72 MB | **Total: 12 files, ≈1.88 million raw records, ≈10.4 GB.** > The C3Med project reports **>1.5M QA pairs** after the annotation pipeline — samples that successfully converged to a correct answer with a valid CoT path are retained for training. --- ## 🪞 Annotation (Re-labeling) Methodology All records are re-labeled by a **three-level pipeline** built around the target LLM itself (backbone: **Qwen3-32B** as the *actor*; **DeepSeek-R1 / Qwen-Plus** as the *critic*): ### ✅ Pipeline A — One-take For **easy questions** that the backbone can answer correctly with **one forward pass**, the model-generated CoT followed by the correct answer is kept directly. ### 🪞 Pipeline B — Self-Reflection For **hard questions** where the backbone initially fails, but an explanation of the ground-truth answer exists in the original dataset: the explanation is shown to the model as a **tip**, and the model is prompted to answer again following the tip. ### 🎭 Pipeline C — Actor-Critic For **hard questions with no available explanation**: the ground-truth answer is disclosed to a powerful **critic model** (with internet access) which is asked to (1) analyze the question and (2) provide an explanation for each option. The explanations are then fed back to the **actor model** as tips for a second attempt. This answering → verifying → reflection loop runs **at most 3 times** until a correct answer is produced. ### 📝 Open-Form Questions For questions without a verifiable ground-truth choice, a critic scores the model-generated answer from **1 to 5** across four dimensions — correctness, completeness, coherence, and helpfulness — with **4.0** as the acceptance threshold. All three pipelines above are applied on this basis as well. **Why re-labeling matters:** CoT paths derived from knowledge graphs or stronger teachers are *heterogeneous* to the target LLM's own reasoning process and degrade SFT effectiveness. By letting the target model reason, fail, reflect, and converge on its own, C3Med-Reasoning produces CoT annotations that are **knowledge-rich and paradigm-aligned** — the key reason C3Med-8B outperforms much larger medical baselines. --- ## 📊 Data Distribution ### 🩺 Task Families | Task Family | Description | | :---------- | :---------- | | 🔤 **Clinical linguistics** | Understanding of clinical terminology & language | | 🧪 **Knowledge-based QA** | Factual medical knowledge question answering | | 💊 **Clinical decision making** | Diagnosis, treatment and management decisions | ### 🏥 Medical Departments The diseases and terminologies in the dataset span **17 medical departments**, including oncology and prevention (肿瘤与预防), internal medicine (内科), surgery (外科) and more. --- ## 🧬 Record Structure Each record typically contains: | Field | Description | | :---- | :---------- | | `question` | The clinical question | | `options` | Answer options (multiple-choice questions) | | `answer` | Ground-truth answer | | `reasoning` | The final re-labeled chain-of-thought path | | `tips` | Tips/explanations used in Pipelines B & C | | `pipeline` | Annotation pipeline used (one-take / self-reflection / actor-critic) | | `source` | Source dataset of the original sample | --- ## 🤖 Trained Model: C3Med This dataset was used to fine-tune the **C3Med** model: <p align="center"> <a href="https://www.modelscope.cn/models/CC-LLM-Research/C3Med"><img src="https://img.shields.io/badge/ModelScope-CC-LLM-Research%2FC3Med-2E86C1" alt="Model"/></a> </p> > **C3Med: Learning LLM with Self-reflection for Chinese-Centered Clinical Medicine** > 8B parameters · Fine-tuned from Qwen3-8B > CMB: **74.12%** avg. (+42.6 pts vs. Baichuan-M2) · Anzhen_Hypertension: **94.84%** avg. · MedS-Ins: **85.87%** 💡 **Tip:** Upload this dataset to your repository or specify it as the training set, then evaluate the resulting model against C3Med on the CMB / MedS-Ins / CL-Bench benchmarks for a direct comparison. --- ## ⬇️ Usage ### ModelScope SDK ```python from modelscope.msdatasets import MsDataset ds = MsDataset.load("CC-LLM-Research/c3med-reasoning") ``` ### Git ```bash git clone https://www.modelscope.cn/datasets/CC-LLM-Research/c3med-reasoning.git ``` --- ## 📚 License & Citation This dataset is released under the **Apache-2.0** license. (Please also respect the licenses and terms of the underlying source datasets.) If you use this dataset, please cite the C3Med model: ```bibtex @misc{c3med2026, title={C3Med: Learning LLM with Self-reflection for Chinese-Centered Clinical Medicine}, author={China Electronics Cloud Technology Co., Ltd.}, howpublished={\url{https://www.modelscope.cn/models/CC-LLM-Research/C3Med}}, year={2026}, note={Dataset: CC-LLM-Research/c3med-reasoning} } ``` --- ## 🔗 Related Resources - **Model:** [CC-LLM-Research/C3Med](https://modelscope.cn/models/CC-LLM-Research/C3Med) - **Dataset:** [CC-LLM-Research/c3med-reasoning](https://modelscope.cn/datasets/CC-LLM-Research/c3med-reasoning) - **Source datasets:** [CMB](https://github.com/FreedomIntelligence/CMB) · [CMExam](https://github.com/williamliwl/CMExam) · [MedS-Ins](https://github.com/WangRongming/MedS-Ins) · [ReasonMed](https://modelscope.cn/datasets/AiMed/ReasonMed) · [Huatuo26M](https://github.com/FreedomIntelligence/Huatuo26M) · [MedQuAD](https://github.com/abachaa/MedQuAD)