CJ-Medical

English

English | 中文

CJ-Medical is the medical-domain checkpoint of Chinese-Jev. Given context, a question, and a set of candidates, it assigns probabilities to the candidates. It supports multiple-choice decisions (choice), proposition judgments (noul), and ordered ratings (score). It does not generate free-form answers.

Training and validation data are available in CJ-Dataset, with held-out evaluation data in CJ-Bench, both under medical/.

Use

The checkpoint uses the Laya decision-model runtime. A standard Transformers text-classification pipeline does not expose its candidate-based interface.

pip install -r requirements.txt

from huggingface_hub import snapshot_download
from laya import Agent

path = snapshot_download("bbldCVer-hf/CJ-Medical")
model = Agent(path, device="cpu")  # use "cuda" on a supported GPU

result = model.system_one(
    state="用户问：体检报告里的这个指标是什么意思？",
    questions={
        "intent": {
            "type": "choice",
            "instructions": "判断用户的问题主要属于哪类意图。",
            "criteria": {
                "A": "检查报告解读",
                "B": "预约挂号",
                "C": "费用咨询",
            },
        }
    },
)
print(result["answers"]["intent"])

state accepts text, an object, or a list. Each question provides its own candidates:

Type	criteria	Returned decision
choice	Object mapping candidate keys to descriptions	Selected key and candidate probabilities
noul	Object with false and true descriptions	noul, the probability of true
score	Ordered list of level descriptions	Level probabilities and expected zero-based level index

Score indices refer to the supplied levels, not a shared medical scale. When reading CJ-Dataset or CJ-Bench, pass only state and questions to the model; keep gold and _meta outside the input.

---

中文

English | 中文

CJ-Medical 是 Chinese-Jev 的医学领域 checkpoint。模型根据上下文、问题和候选项输出候选概率，支持选择题（choice）、命题判断（noul）和有序评分（score），不生成开放式回答。

训练与验证数据见 CJ-Dataset，独立测试数据见 CJ-Bench，均位于 medical/ 目录。

使用

运行上方示例即可加载模型。checkpoint 使用 Laya 的决策模型接口，不能直接通过普通 Transformers 文本分类 pipeline 调用候选判断功能。

state 可为文本、对象或列表。每个问题分别定义候选：

类型	criteria	输出
choice	候选键与描述组成的对象	选中的键及候选概率
noul	包含 false、true 描述的对象	noul 字段，即 true 的概率
score	有序等级描述列表	等级概率及从 0 开始的等级索引期望

Score 的索引对应题目给出的等级，不代表统一医学量表。读取 CJ-Dataset 或 CJ-Bench 时，只把 state、questions 传给模型，gold 和 _meta 留在监督与分析侧。