CJ-Bench

English

English | 中文

CJ-Bench evaluates Chinese conditional decision models on 307,900 test decisions across general language, medicine, law, and finance. Each case provides context, a question, and finite candidates; a model returns a probability distribution over those candidates. Training and validation data are in CJ-Dataset.

Composition

Domain	Choice	Noul	Score	Total	ACC/ECE eligible
General	33,334	33,333	33,333	100,000	96,922
Medical	98,543	98,544	2,913	200,000	200,000
Legal	1,500	1,500	1,300	4,300	4,300
Finance	1,200	1,200	1,200	3,600	3,600
Total	134,577	134,577	38,746	307,900	304,822

Input limits are 1,024 tokens for General and Finance, 2,048 for Medical, and 8,192 for Legal, measured with the Laya Multilingual tokenizer and input compiler. Medical also has a 256-token decision-header limit. Counts refer to decisions; several may share one source question or document.

Sources and construction

General covers reading and reasoning, inference, semantic matching, sentiment, reviews, categories, and intent. Medical covers examinations, clinical and dialogue annotations, terminology, and query matching. Legal covers causes of action, claim outcomes, case relevance, norm conflicts, privacy policies, and explicit rules. Finance covers examinations, entity relations, dialogue strategies, answer relevance, readability, and explicit rules. The source inventory lists upstream releases and retained counts.

Source annotations were converted to candidate distributions. Deterministic tasks use explicit rules with programmatically checked answers. General is a stratified 100K subset of its held-out pool; Medical is a 200K subset of its original integrated test partition. Derived views stay with their designated parent question, case, or conversation. LCR-CN retains an official clause-level evaluation protocol: different clauses from one regulation may appear across splits.

Format

Each gzip-compressed JSONL line contains one decision. This illustrative example is not a benchmark item:

{
  "id": "illustration-001",
  "state": "会议改到明天下午三点，地点仍是二楼会议室。",
  "questions": {
    "q": {
      "type": "choice",
      "instructions": "这条消息通知了什么变化？",
      "criteria": {"A": "会议时间", "B": "会议地点", "C": "参会人员"}
    }
  },
  "gold": {
    "q": {"type": "choice", "probabilities": {"A": 1.0, "B": 0.0, "C": 0.0}}
  }
}

Only state and questions are model inputs. gold holds the target distribution; _meta, when present, records source, grouping, and task information. state may be a string, object, or list. Match questions and targets by key; the key is not always q.

Type	Candidates and target keys
choice	Named alternatives in a criteria object; probability keys match its keys.
noul	A proposition judgment, with probability keys "false" and "true", in that order.
score	An ordered criteria list; probability keys "0", "1", … index its levels.

Score indices are positions, not a universal rating scale. Read the criterion text and any source-scale mapping in _meta. A multi-answer exam converted to Noul asks whether an option belongs to the original answer key, including when the question asks for an incorrect statement.

Download

Files are stored as <domain>/test.cases.jsonl.gz. Install huggingface_hub and authenticate with hf auth login or HF_TOKEN while repository access is restricted. Read the gzip stream directly:

import gzip
import json
from huggingface_hub import hf_hub_download

domain = "general"  # general, medical, legal, finance
split = "test"
path = hf_hub_download(
    repo_id="bbldCVer-hf/CJ-Bench",
    repo_type="dataset",
    filename=f"{domain}/{split}.cases.jsonl.gz",
)
with gzip.open(path, "rt", encoding="utf-8") as stream:
    for line in stream:
        case = json.loads(line)
        model_input = {k: case[k] for k in ("state", "questions")}
        target = case["gold"]

Evaluation

Report results separately for General, Medical, Legal, and Finance, using the same retained context, candidates, and eligible cases for every model.

• ACC: the fraction of eligible decisions whose highest-probability candidate matches the gold label.
• ECE: the gap between top-candidate confidence and accuracy, using 15 equal-width bins weighted by their sample counts.
• Latency: median end-to-end time per decision after warm-up. The report uses one NVIDIA H200 and batch size 1, including input preparation and inference through final candidate probabilities. Hosted API latency is measured at the client for one decision per request, including the network round trip.

ACC and ECE include hard, single-label targets of all three types, including Score. General's 3,078 soft-target cases are excluded, leaving 96,922 eligible General cases. All specialist cases have hard targets. The report retains each checkpoint's native probability transformations and applies no additional target-dataset calibration. Test labels must not be used to choose thresholds or fit calibration.

Scope and limitations

The selected distributions reflect benchmark design rather than real-world request frequencies. Medical Score measures only QTR query–title relevance. Legal includes long inputs, and rule-based synthetic cases contribute 1,498 of its 4,300 decisions; report length filtering or subset selection explicitly. Known parent identifiers can support grouped uncertainty estimates, since rows are not all independent. Cross-domain overlap and model-pretraining contamination have not been exhaustively audited.

Use and licenses

This release is intended for research on Chinese decision models. Annotations retain their upstream meaning: answer matching, for example, does not establish medical correctness, and a historical legal label does not establish current legal advice. Some labels are model-generated or come from mixed annotation pipelines, including AppealCase and RoleCS. The collection has not received a new, exhaustive human review or a corpus-wide privacy audit.

Licenses apply by source, as recorded in _meta and the source inventory. ClaimGen-CN and CFLUE carry CC BY-NC-SA 4.0 terms; AppealCase carries CC BY-NC 4.0 terms. Several General and Medical sources have research-use restrictions, and some Medical releases lack a confirmed independent data license. license: other reflects these differences and grants no additional rights; commercial use or redistribution requires checking the applicable source terms.

Version, citation, and contact

Files were copied without recompression from chinese-jev revision a3963a61f71e00051a80721e587d720d8926e8e7. MANIFEST.json records counts, sizes, and checksums. Pin this repository's revision when reporting experiments.

This release accompanies Chinese-Jev: Bringing System One Model to Chinese-Language Tasks. To cite the data, use the entry below. Report corrections in Discussions, with the revision and case ID.

@misc{cjbench2026,
  author = {{Chinese-Jev Contributors}},
  title = {CJ-Bench: Chinese-Jev Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/bbldCVer-hf/CJ-Bench}
}

---

中文

English | 中文

CJ-Bench 包含 307,900 条测试 decisions，用于评测中文条件化决策模型，覆盖通用、医学、法律和金融四个领域。每条记录给出上下文、问题及有限候选，模型输出候选概率分布。训练与验证数据见 CJ-Dataset。

数据组成

领域	Choice	Noul	Score	总数	ACC/ECE 有效样本
General	33,334	33,333	33,333	100,000	96,922
Medical	98,543	98,544	2,913	200,000	200,000
Legal	1,500	1,500	1,300	4,300	4,300
Finance	1,200	1,200	1,200	3,600	3,600
合计	134,577	134,577	38,746	307,900	304,822

按 Laya Multilingual tokenizer 与输入编译器计数，General、Finance 的输入上限为 1,024 tokens，Medical 为 2,048，Legal 为 8,192；Medical 另有 256-token 的题头上限。表中计数单位为 decision，同一原题或文档可能产生多条。

来源与构造

General 覆盖阅读推理、文本推断、语义匹配、情感、评论、类别与意图；Medical 覆盖考试、临床及对话标注、术语与查询匹配；Legal 覆盖案由、诉请结果、类案相关性、规范冲突、隐私政策和明示规则；Finance 覆盖考试、实体关系、对话策略、回答相关性、可读性和明示规则。来源清单 列出上游入口及保留数量。

原标注转换为候选分布；确定性任务采用题内明示规则，并由程序核验答案。General 是从留出池分层选取的 100K 子集；Medical 是从原医学整合测试划分中选取的 200K 子集。派生视图随指定的原问题、案件或对话分组。LCR-CN 保留官方条款级评测协议，同一规范的不同条款可以出现在不同划分中。

数据格式

文件采用 gzip 压缩的 JSONL，每行一个 decision。以下为格式示例，不是评测题：

{
  "id": "illustration-001",
  "state": "会议改到明天下午三点，地点仍是二楼会议室。",
  "questions": {
    "q": {
      "type": "choice",
      "instructions": "这条消息通知了什么变化？",
      "criteria": {"A": "会议时间", "B": "会议地点", "C": "参会人员"}
    }
  },
  "gold": {
    "q": {"type": "choice", "probabilities": {"A": 1.0, "B": 0.0, "C": 0.0}}
  }
}

模型只读取 state 和 questions。gold 保存目标分布；_meta（如有）保存来源、分组和任务信息。state 可以是字符串、对象或列表。问题与目标按键对应，不能假定键名总是 q。

类型	候选与概率键
choice	criteria 对象中的命名候选；概率键与候选键一致。
noul	命题判断，概率键按 "false"、"true" 排列。
score	criteria 有序列表；概率键 "0"、"1" 等对应等级位置。

Score 索引是位置，不是统一评分尺度，应结合候选描述及 _meta 中的原量表映射解释。多答案考试转换成 Noul 后，判断的是该选项是否属于原答键；“选错误项”等否定题仍沿用这一含义。

下载

文件路径为 <domain>/test.cases.jsonl.gz。安装 huggingface_hub；仓库访问受限期间，请通过 hf auth login 或 HF_TOKEN 认证。可直接读取 gzip 流：

import gzip
import json
from huggingface_hub import hf_hub_download

domain = "general"  # general, medical, legal, finance
split = "test"
path = hf_hub_download(
    repo_id="bbldCVer-hf/CJ-Bench",
    repo_type="dataset",
    filename=f"{domain}/{split}.cases.jsonl.gz",
)
with gzip.open(path, "rt", encoding="utf-8") as stream:
    for line in stream:
        case = json.loads(line)
        model_input = {k: case[k] for k in ("state", "questions")}
        target = case["gold"]

评测

General、Medical、Legal、Finance 分别报告结果。各模型使用相同的保留上下文、候选及有效样本。

• ACC： 最高预测概率候选与金标一致的有效 decision 比例。
• ECE： 使用 15 个等宽置信度区间，按区间样本数加权，衡量最高概率候选的置信度与准确率之差。
• 延迟： 预热后的单条 decision 端到端耗时中位数。报告使用一张 NVIDIA H200、batch size 1，包含输入准备及推理至最终候选概率的全部步骤。托管 API 每次请求一条 decision，记录含网络往返的客户端耗时。

ACC、ECE 使用三种题型中的硬单标签样本，包括硬标签 Score。General 的 3,078 条软目标样本不参与这两项指标，剩余 96,922 条有效样本；三个垂域均为硬目标。报告保留各 checkpoint 原有的概率变换，不额外在目标数据集上进行校准。测试标签不得用于选择阈值或拟合校准参数。

适用范围与局限

选样分布反映评测设计，不代表现实请求频率。Medical Score 仅衡量 QTR 查询—标题相关性。Legal 包含长输入，且 4,300 条中有 1,498 条为程序规则题；额外长度过滤或子集选择应明确报告。多条记录可能共享母样本，估计不确定性时可使用已有母组标识。跨领域重复及模型预训练污染尚未完成穷尽核查。

使用与许可

本数据用于中文决策模型研究，标签沿用原任务含义。例如，问答匹配不等于医学正确性，历史法律标签不等于现行法律意见。部分标签来自模型或混合标注流程，包括 AppealCase 和 RoleCS。整合后未进行新一轮逐条人工复核，也未完成覆盖全部语料的隐私审计。

许可按来源适用，见 _meta 与来源清单。ClaimGen-CN、CFLUE 使用 CC BY-NC-SA 4.0，AppealCase 使用 CC BY-NC 4.0。部分通用和医学来源限于研究用途；部分医学来源尚未确认独立的数据许可。license: other 反映这些差异，不授予额外权利；商业使用或再分发前须核对对应来源条款。

版本、引用与反馈

数据文件来自 chinese-jev 的固定版本 a3963a61f71e00051a80721e587d720d8926e8e7，未重新压缩。MANIFEST.json 记录各文件的数量、大小和校验值。复现实验时请固定本仓库 revision。

本数据配套技术报告 Chinese-Jev: Bringing System One Model to Chinese-Language Tasks。引用数据时可使用下方条目；问题或更正请提交至 Discussions，注明 revision 与 case ID。

@misc{cjbench2026,
  author = {{Chinese-Jev Contributors}},
  title = {CJ-Bench: Chinese-Jev Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/bbldCVer-hf/CJ-Bench}
}