---
language:
- zh
pretty_name: "CJ-Dataset: Chinese-Jev Training Dataset"
license: other
license_name: source-specific
license_link: https://huggingface.co/datasets/bbldCVer-hf/CJ-Dataset#use-and-licenses
task_categories:
- text-classification
- question-answering
size_categories:
- 10M<n<100M
tags:
- chinese
- decision-making
- calibration
- jev
- medical
- legal
- finance
---

# CJ-Dataset

## English

**English** | [中文](#中文)

CJ-Dataset provides **13,048,000 training decisions** and **134,318 validation decisions** for Chinese conditional decision models. It combines general-language tasks with medical, legal, and financial specialization in a shared context–question–candidates format. Final evaluation sets are released separately in [CJ-Bench](https://huggingface.co/datasets/bbldCVer-hf/CJ-Bench).

### Composition

| Domain | Train Choice | Train Noul | Train Score | Train total | Validation |
|---|---:|---:|---:|---:|---:|
| General | 3,333,334 | 3,333,333 | 3,333,333 | 10,000,000 | 30,000 |
| Medical | 1,194,496 | 1,792,223 | 13,281 | 3,000,000 | 100,000 |
| Legal | 8,000 | 8,000 | 8,000 | 24,000 | 2,518 |
| Finance | 8,000 | 8,000 | 8,000 | 24,000 | 1,800 |
| **Total** | **4,543,830** | **5,141,556** | **3,362,614** | **13,048,000** | **134,318** |

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 combines 20 public sources with a 10,700-decision programmatic supplement. Tasks include reading and reasoning, inference, matching, sentiment, review ratings, categories, and intent. Medical draws from 30 source families spanning examinations, clinical text, dialogue, terminology, and query matching. Legal uses ClaimGen-CN, AppealCase, LeCaRDv2, eCAIL, LCR-CN, CAPP-130, and explicit rules. Finance uses CFLUE, FinRE, RoleCS, FinTruthQA, and explicit rules. See the [source inventory](SOURCES.md) for contributions, upstream links, and recorded license profiles.

Conversion retains the meaning of each source label. Single-answer questions become Choice; binary propositions and option membership become Noul; ordered ratings become Score. Programmatic supplements use explicit rules and executable answer checks. Grouping, duplicate and conflict checks, and token-length validation precede selection. Length checks apply to the task's retained context and candidates; for example, some RoleCS tasks explicitly use recent complete dialogue turns.

Derived views are assigned with their designated parent material. Original held-out partitions take priority where available. LCR-CN uses an official clause-level protocol that permits distinct clauses from the same regulation across splits. The Medical-3M selection keeps the original integrated train, validation, and test partitions separate; multi-answer questions retain complete option bundles after length checks. General's released 30K validation file is a compact subset of its larger development pool. Calibration sets are not included here.

### Format

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

```json
{
  "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

Each domain has `train.cases.jsonl.gz` and `validation.cases.jsonl.gz`. Install `huggingface_hub` and authenticate with `hf auth login` or `HF_TOKEN` while repository access is restricted. The files can be read without an extra decompressed copy:

```python
import gzip
import json
from huggingface_hub import hf_hub_download

domain = "general"  # general, medical, legal, finance
split = "train"  # train or validation
path = hf_hub_download(
    repo_id="bbldCVer-hf/CJ-Dataset",
    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"]
```

### Training and limitations

The report first trains on General, then initializes three independent domain specialists from that checkpoint. Validation supports model selection; CJ-Bench is reserved for final evaluation. Its ACC/ECE protocol uses hard, single-label cases and excludes soft targets.

Medical remains imbalanced after selection: cMedQA2 contributes 1,486,404 training decisions, and all 13,281 Score decisions use the QTR relevance scale. Sampling can change their exposure, not the number or variety of independent annotations. Legal and Finance contain 2,200 and 5,000 programmatic training decisions, respectively, governed by fictional rules stated in the input. They measure rule application rather than observed professional outcomes. Coverage remains uneven across source tasks, and 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](SOURCES.md). 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](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](https://huggingface.co/datasets/bbldCVer-hf/CJ-Dataset/discussions), with the revision and case ID.

```bibtex
@misc{cjdataset2026,
  author = {{Chinese-Jev Contributors}},
  title = {CJ-Dataset: Chinese-Jev Training Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/bbldCVer-hf/CJ-Dataset}
}
```

---

## 中文

[English](#english) | **中文**

CJ-Dataset 为中文条件化决策模型提供 **13,048,000 条训练 decisions** 和 **134,318 条验证 decisions**。数据包含通用、医学、法律和金融任务，共享“上下文—问题—候选”格式。最终评测集单独发布于 [CJ-Bench](https://huggingface.co/datasets/bbldCVer-hf/CJ-Bench)。

### 数据组成

| 领域 | Train Choice | Train Noul | Train Score | Train 总数 | Validation |
|---|---:|---:|---:|---:|---:|
| General | 3,333,334 | 3,333,333 | 3,333,333 | 10,000,000 | 30,000 |
| Medical | 1,194,496 | 1,792,223 | 13,281 | 3,000,000 | 100,000 |
| Legal | 8,000 | 8,000 | 8,000 | 24,000 | 2,518 |
| Finance | 8,000 | 8,000 | 8,000 | 24,000 | 1,800 |
| **合计** | **4,543,830** | **5,141,556** | **3,362,614** | **13,048,000** | **134,318** |

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 包含 20 个公开来源及 10,700 条程序构造题，覆盖阅读推理、文本推断、匹配、情感、评论评分、类别与意图。Medical 来自 30 个来源家族，涵盖考试、临床文本、对话、术语与查询匹配。Legal 使用 ClaimGen-CN、AppealCase、LeCaRDv2、eCAIL、LCR-CN、CAPP-130 及明示规则；Finance 使用 CFLUE、FinRE、RoleCS、FinTruthQA 及明示规则。各源贡献、上游入口与记录的许可条件见[来源清单](SOURCES.md)。

转换保留原标签含义：单答案题转为 Choice，二元命题和选项成员判断转为 Noul，有序等级转为 Score。程序补充题使用明示规则及可执行的答案检查。分组、重复与冲突检查、分词长度验证先于选样。长度检查针对任务定义中保留的上下文及候选；例如，部分 RoleCS 任务明确只使用最近的完整对话轮次。

派生视图随指定母材料划分，并优先保护已有官方留出。LCR-CN 采用官方条款级协议，允许同一规范的不同条款跨划分出现。Medical-3M 保持原医学整合池的 train、validation、test 隔离；多答案题经过长度检查后保留完整选项包。General 发布的 30K validation 是较大开发池的紧凑子集。本仓库不包含 calibration。

### 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：

```json
{
  "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

每个领域包含 `train.cases.jsonl.gz` 与 `validation.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。读取时无需额外保存解压副本：

```python
import gzip
import json
from huggingface_hub import hf_hub_download

domain = "general"  # general, medical, legal, finance
split = "train"  # train or validation
path = hf_hub_download(
    repo_id="bbldCVer-hf/CJ-Dataset",
    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"]
```

### 训练与局限

技术报告先使用 General 训练，再从该 checkpoint 分别初始化三个领域模型。Validation 用于模型选择，CJ-Bench 用于最终评测；其中 ACC/ECE 使用硬单标签样本并排除软目标。

Medical 选样后仍存在不均衡：cMedQA2 贡献 1,486,404 条训练 decisions，13,281 条 Score 全部来自 QTR 相关性量表。采样可以调整曝光，不能增加独立标注的数量或种类。Legal、Finance 分别有 2,200、5,000 条程序训练题，依据输入中的虚构规则监督，衡量规则应用而非真实业务结果。各任务覆盖仍不均匀，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](SOURCES.md)。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](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](https://huggingface.co/datasets/bbldCVer-hf/CJ-Dataset/discussions)，注明 revision 与 case ID。

```bibtex
@misc{cjdataset2026,
  author = {{Chinese-Jev Contributors}},
  title = {CJ-Dataset: Chinese-Jev Training Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/bbldCVer-hf/CJ-Dataset}
}
```