TraceMed

TraceMed

TraceMed 是采用 Qwen3 架构的文本生成模型。本仓库提供训练后的模型权重、配置、分词器和聊天模板，可通过 Hugging Face Transformers 加载使用。

模型信息

能力展示 / Capability Showcase

以下截图来自我们的内部对话测试界面，左栏为 TraceMed，右栏为同尺寸基座模型 Qwen3-4B（开启思考模式），输入完全相同。回答中的角标（如 [1]）可点击，直接定位到被引用的原文段落。

The screenshots below are taken from our internal chat-testing UI. The left pane is TraceMed; the right pane is the same-size base model Qwen3-4B (thinking mode on), given the identical input. Superscript markers (e.g. [1]) in our model's answers are clickable and jump to the exact source passage being cited.

主要能力 / Key capabilities

---

示例 1：湿温病辨证 / Example 1: Pattern Differentiation of Damp-Warmth

输入 / Input

患者身热不扬，午后热甚，头痛恶寒，身重疼痛，胸闷不饥，舌白不渴，脉弦细而濡。请辨证并给出治法和代表方，引用医学依据。

A patient presents with a low-grade fever that does not feel hot to the touch and worsens in the afternoon, headache with aversion to cold, a heavy and aching body, chest oppression with no appetite, a white tongue coating without thirst, and a wiry, thin, soggy pulse. Please identify the pattern, give the treatment principle and a representative formula, and cite medical evidence.

输出（节选）/ Output (excerpt)

辨证分析
根据患者临床表现，辨证为湿温病，证属湿重于热（或湿温初起，邪遏卫气，湿重于热）。
1. 症状辨析：
 · 身热不扬，午后热甚：湿性黏滞，遏阻热邪，故身热不扬；午后阳气渐衰，湿邪得阴助而势张，故午后热势加重[1][2]。
 · 头痛恶寒，身重疼痛：湿邪郁遏卫阳，清阳不升，故头痛；湿性重浊，阻滞经络，故身重疼痛[3][4]。
 · 胸闷不饥，舌白不渴：湿阻中焦，气机不畅，故胸闷不饥；湿邪内蕴，未化燥伤津，故舌苔白腻、口不渴[5][6]。
2. 病机概要：
 本病病机为湿邪偏盛，热邪被湿所裹而不得外达，湿遏热伏，困阻中焦，气机不畅[8][9]。病位主要在上焦及中焦，涉及卫分与气分[10][11]。
……（以下略）

Pattern Analysis
Based on the clinical presentation, this is damp-warmth disease with dampness predominating over heat (early-stage damp-warmth, pathogen obstructing the defensive and qi levels).
1. Symptom analysis:
 · Low-grade fever worse in the afternoon: dampness is sticky and traps heat, so the fever does not surface; in the afternoon yang qi declines and dampness, aided by yin, intensifies, so the fever worsens [1][2].
 · Headache, aversion to cold, heavy aching body: dampness constrains defensive yang so clear yang cannot ascend, causing headache; dampness is heavy and turbid, obstructing the channels, causing heaviness and pain [3][4].
 · Chest oppression, no appetite, white coating, no thirst: dampness obstructs the middle burner and impedes qi movement; dampness has not yet transformed into dryness to damage fluids, hence the white greasy coating and absence of thirst [5][6].
2. Pathogenesis:
 Dampness predominates and envelops heat so it cannot vent outward; heat is hidden beneath dampness, obstructing the middle burner and qi movement [8][9]. The disease is located mainly in the upper and middle burners, involving the defensive and qi levels [10][11].
… (truncated)

---

示例 2：方剂对比与引用溯源 / Example 2: Formula Comparison with Citation Tracing

输入 / Input

麻黄汤与桂枝汤在组成、功效、适应证上有何异同？

What are the similarities and differences between Mahuang Decoction and Guizhi Decoction in composition, function and indications?

输出（节选）/ Output (excerpt)

麻黄汤与桂枝汤均含有桂枝和甘草，体现了方剂组方中"药对"的基础作用[1]。两者在配伍法度上均遵循阴阳相对的原则，如桂枝配麻黄、桂枝配芍药等，体现了严谨的组方逻辑[2]。
在功效方面，麻黄汤具有辛温发汗、宣肺平喘的功效[3][4]。桂枝汤则具有辛温解肌、调和营卫的功效[5][6][7][8][9]。麻黄汤与桂枝汤的共同功效包括解表祛邪、调和营卫、宣通经脉[10]。……
在适应证上，两者存在显著差异。麻黄汤与桂枝汤均具有辛温解表的功效，但麻黄汤发汗力较强，适用于表实无汗证……

Both Mahuang Decoction and Guizhi Decoction contain Guizhi (cinnamon twig) and Gancao (licorice), reflecting the foundational role of herb pairs in formula construction [1]. Both follow the principle of complementary yin–yang pairing, e.g. Guizhi with Mahuang and Guizhi with Baishao, showing rigorous compositional logic [2].
In function, Mahuang Decoction is pungent-warm, induces sweating, and diffuses the lung to relieve wheezing [3][4]. Guizhi Decoction is pungent-warm, releases the muscle layer and harmonizes the nutritive and defensive qi [5][6][7][8][9]. …
In indications they differ markedly: both release the exterior with pungent-warm herbs, but Mahuang Decoction has a stronger diaphoretic effect and suits exterior-excess patterns without sweating …

引用溯源 / Citation tracing：点击角标 [3]，右侧面板显示「精确命中」的原文出处与段落： Clicking marker [3] opens the side panel showing an exact match to its source passage:

出处 / Source：伤寒论临床运用 › 第二章 太阳病辨证论治 › 第二节 太阳病本证证治 › 二、太阳伤寒表实证 › （一）麻黄汤证 › 【方药精解】 › 宣肺平喘
引用句 / Cited sentence：在功效方面，麻黄汤具有辛温发汗、宣肺平喘的功效
原文 / Passage：宣肺平喘：麻黄又有宣肺平喘之功，杏仁，宣肺降气，助麻黄平喘。……

---

示例 3：医学考试选择题 / Example 3: Medical Exam Multiple-Choice Question

输入 / Input

治疗风寒表实证的代表方是：A. 桂枝汤 B. 麻黄汤 C. 银翘散 D. 小青龙汤 E. 九味羌活汤。请给出答案并说明理由。

Which is the representative formula for the wind-cold exterior-excess pattern? A. Guizhi Decoction B. Mahuang Decoction C. Yinqiao Powder D. Xiao Qinglong Decoction E. Jiuwei Qianghuo Decoction. Give the answer and explain why.

输出 / Output

{
  "answer": "B",
  "rationale": "风寒表实证（又称风寒束肺证或太阳伤寒证）的病机为风寒外束肌表，导致卫阳被遏、营阴郁滞、肺气失宣[1][2][3]。其治疗原则是疏风散寒、宣肺止咳[4]。麻黄汤出自《伤寒论》，是治疗风寒表实证的代表方剂[5]，具有发汗解表、宣肺平喘的功效[6][7][8]。方中麻黄为君药，发汗散寒、宣肺平喘[9]；桂枝为臣药，助麻黄发汗解肌[10]；杏仁为佐药，降利肺气以助平喘[11]。因此，治疗风寒表实证的代表方是麻黄汤[12]。桂枝汤主要用于风寒表虚证[13]，银翘散用于风热表证，小青龙汤用于外寒内饮证[14]，均非风寒表实证的首选代表方。"
}

{
  "answer": "B",
  "rationale": "The wind-cold exterior-excess pattern (also called Taiyang cold damage) arises when wind-cold fetters the exterior, constraining defensive yang, stagnating nutritive yin and impairing the lung's diffusing function [1][2][3]. Treatment is to disperse wind-cold and diffuse the lung [4]. Mahuang Decoction, from the Shanghan Lun, is the representative formula for this pattern [5]; it induces sweating to release the exterior and diffuses the lung to relieve wheezing [6][7][8]. Mahuang is the sovereign herb, inducing sweating and dispersing cold [9]; Guizhi is the minister, assisting sweating and releasing the muscle layer [10]; Xingren is the assistant, directing lung qi downward to aid in relieving wheezing [11]. Hence the answer is Mahuang Decoction [12]. Guizhi Decoction treats the exterior-deficiency pattern [13], Yinqiao Powder treats wind-heat exterior patterns, and Xiao Qinglong Decoction treats external cold with internal fluid retention [14]; none is the first-choice formula here."
}

---

免责声明 / Disclaimer：本模型仅用于研究与教学，输出内容不构成医疗建议，不能替代执业医师的诊断与治疗。 This model is intended for research and educational purposes only. Its outputs do not constitute medical advice and must not replace diagnosis or treatment by a licensed physician.

快速使用

使用 uv 安装依赖：

uv pip install -U torch transformers accelerate huggingface_hub

如果仓库为私有，请先使用有访问权限的账号登录：

uv tool run --from huggingface_hub hf auth login

在模型权重上传完成后运行以下示例。示例需要足够内存或显存，具体需求取决于设备和输入长度。

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "GancaoDoctorAI/TraceMed"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)
model.eval()

messages = [{"role": "user", "content": "你好，请介绍一下你自己。"}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
reply = outputs[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(reply, skip_special_tokens=True))

文件说明

• model.safetensors：模型权重。
• config.json：模型结构配置。
• generation_config.json：默认生成配置。
• tokenizer.json、tokenizer_config.json：分词器及配置。
• chat_template.jinja：对话输入模板。

补充信息

基础模型来源、训练数据、评测结果与许可证信息待补充。