Keep K3 suite selection and report-schema scoring in bash, merge K3/vision dataset_args into dpv4 yamls, and pin EvalScope at 735d920ee911 with local patches. Co-authored-by: Cursor <cursoragent@cursor.com>
163 lines
10 KiB
Markdown
163 lines
10 KiB
Markdown
# PLawBench
|
||
|
||
|
||
## Overview
|
||
|
||
PLawBench is a rubric-based benchmark that evaluates large language models on real-world Chinese legal practice.
|
||
It mirrors the workflow of a practising lawyer across three hierarchical levels: eliciting facts during a public
|
||
legal consultation, analysing a case with structured legal reasoning, and drafting professional legal documents.
|
||
Every item ships with a rubric annotated by legal experts, and grading is performed by an LLM judge against that
|
||
rubric rather than against a single reference answer.
|
||
|
||
## Task Description
|
||
|
||
- **Task Type**: Open-ended Chinese legal generation graded with expert rubrics
|
||
- **Input**: A client statement, or a case description plus a legal question
|
||
- **Output**: A question list, a structured case analysis, or a full legal document
|
||
- **Domain**: Chinese legal practice (personal affairs, marriage and family, corporate governance, intellectual
|
||
property, criminal and civil litigation, cross-border matters, labour, environmental safety, and more)
|
||
|
||
## Key Features
|
||
|
||
- 280 samples split into four subsets, one per PLawBench task:
|
||
- `case_analysis` (250): case analysis scored on four dimensions — conclusion, case facts, reasoning, and cited
|
||
statutes. Answers must follow the 【结论】/【案件事实】/【推理过程】/【法条依据】 structure.
|
||
- `legal_consultation` (18): the model plays a lawyer and must produce 10-25 verifiable follow-up questions that
|
||
surface the facts the client omitted or distorted.
|
||
- `plaintiff_statement` (6): drafting a statement of complaint from the client's account.
|
||
- `defendant_statement` (6): drafting a statement of defense from the client's account and the opposing complaint.
|
||
- Client statements are deliberately vague, emotional, or misleading, so models must detect traps instead of
|
||
restating the client's claims.
|
||
- Task prompts and judge prompts are ported verbatim from the official release, and the `case_analysis` rubric
|
||
retains its per-dimension point allocation.
|
||
|
||
## Evaluation Notes
|
||
|
||
- Requires an LLM judge: set `judge.strategy='llm'` (or `'auto'`, which enables the judge for this benchmark)
|
||
and provide `judge.models`. `judge.strategy='rule'` is not supported.
|
||
- Metrics are point ratios in `[0, 1]`. `acc` is reported for every subset; `case_analysis` additionally reports
|
||
`conclusion_acc`, `fact_acc`, `reasoning_acc`, and `law_acc`. These map one-to-one onto the official leaderboard
|
||
columns: `legal_consultation` is Task1, `case_analysis` is Task2-Avg with its four dimensions, and the two
|
||
drafting subsets are Task3-Plaintiff and Task3-Defendant.
|
||
- Compare per-subset scores, not the `OVERALL` row. `OVERALL` is a per-sample mean, so `case_analysis` dominates it
|
||
(250 of 280 samples). The paper's `Overall` column is an equal-weighted mean of the three task scores, which
|
||
matches its published table far more closely (mean absolute error 0.72 versus 2.87 for a sample-weighted mean,
|
||
fitted across the 24 models in the official ranking).
|
||
- Rubric point totals come from the dataset, not from the judge output, and awarded points are clamped into
|
||
`[0, max_points]`, so a judge that mis-reports the denominator cannot distort the score.
|
||
- The judge output template for `case_analysis` is repaired relative to the official script, which ships malformed
|
||
JSON and pins the conclusion section to zero points; every section is graded on its rubric allocation here.
|
||
- The judge model's transport retry policy is configured through its `generation_config`. A reply that still
|
||
fails the output contract is unavailable and excluded rather than silently scored as zero.
|
||
- Case-analysis judging returns a long per-item breakdown. Give the judge a generous `max_tokens`
|
||
(for example 8192) in `judge.models[].generation_config`.
|
||
- The drafting subsets ask for a 2,500-3,000 character legal document, so the evaluated model also needs a generous
|
||
`generation_config.max_tokens`. A truncated filing is graded as an incomplete document and scores near zero, which
|
||
depresses Task3 for reasons unrelated to legal ability.
|
||
|
||
Resources: [GitHub](https://github.com/skylenage/PLawbench) |
|
||
[Dataset](https://modelscope.cn/datasets/evalscope/PLawBench)
|
||
|
||
|
||
## Properties
|
||
|
||
| Property | Value |
|
||
|----------|-------|
|
||
| **Benchmark Name** | `plawbench` |
|
||
| **Dataset ID** | [evalscope/PLawBench](https://modelscope.cn/datasets/evalscope/PLawBench/summary) |
|
||
| **Paper** | [Paper](https://github.com/skylenage/PLawbench) |
|
||
| **Tags** | `Chinese`, `Knowledge`, `QA`, `Reasoning` |
|
||
| **Metrics** | `accuracy`, `conclusion_acc`, `fact_acc`, `reasoning_acc`, `law_acc` |
|
||
| **Default Shots** | 0-shot |
|
||
| **Evaluation Split** | `test` |
|
||
|
||
|
||
## Data Statistics
|
||
|
||
| Metric | Value |
|
||
|--------|-------|
|
||
| Total Samples | 280 |
|
||
| Prompt Length (Mean) | 2669.88 chars |
|
||
| Prompt Length (Min/Max) | 1267 / 5890 chars |
|
||
|
||
**Per-Subset Statistics:**
|
||
|
||
| Subset | Samples | Prompt Mean | Prompt Min | Prompt Max |
|
||
|--------|---------|-------------|------------|------------|
|
||
| `case_analysis` | 250 | 2720.18 | 2137 | 4873 |
|
||
| `legal_consultation` | 18 | 1851.94 | 1512 | 2611 |
|
||
| `plaintiff_statement` | 6 | 1494.67 | 1267 | 2050 |
|
||
| `defendant_statement` | 6 | 4203 | 2794 | 5890 |
|
||
|
||
## Sample Example
|
||
|
||
**Subset**: `case_analysis`
|
||
|
||
```json
|
||
{
|
||
"input": [
|
||
{
|
||
"id": "06c8a1d7",
|
||
"content": "\n## 角色\n\n你是一名具有十年以上执业经验的法律实务专家,精通中国现行法律法规与司法实践。你擅长将复杂的法律问题分解为清晰的逻辑模块,并严格依据“结论先行、事实为重、推理严密、依据支撑”的专业风格进行解答。\n\n## 核心要求\n\n1. 严格顺序:回答必须按照以下四部分顺序展开,并使用对应标题:\n【结论】\n【案件事实】\n【推理过程】\n【法条依据】\n2. 内容规范:\n结论:直接、明确,针对提问的核心争议点给出肯定或否定的判断。\n案件事实:基于用户提供的案情,简明、客观地摘录与法律判断 ... [TRUNCATED 1912 chars] ... 并赔偿精神损害抚慰金。庭审中查明,某摄影服务公司已完成除摄像外的其他服务项目;某文化传媒公司系独立法人,其工作人员在操作设备时存在重大过失。另查,某甲在签订合同时未特别声明婚礼录像的重要性,但合同附件中列有\"全程跟拍记录\"服务项目。某摄影服务公司辩称其仅需承担合同违约责任,精神损害赔偿缺乏依据。某文化传媒公司以非合同相对方为由拒绝承担责任。\n\n## 问题\n以【结论 + 案情简述 + 分析过程+依据法条】的逻辑回答以下问题:在上述案例中,某甲能否向某摄影服务公司主张精神损害赔偿?\n"
|
||
}
|
||
],
|
||
"target": "",
|
||
"id": 0,
|
||
"group_id": 0,
|
||
"subset_key": "case_analysis",
|
||
"metadata": {
|
||
"id": "case_analysis-1",
|
||
"task": "case_analysis",
|
||
"judge_type": "case_analysis",
|
||
"category": "个人生活",
|
||
"rubrics": "[{\"criterion\": \"【结论得分】\\n(+5分) 某甲有权向某摄影服务公司主张精神损害赔偿。\", \"points\": \"5\", \"tags\": \"结论得分\"}, {\"criterion\": \"【案情简述得分】\\n(+5分) 某甲与某摄影服务公司签订《婚庆服务合同》,并支付全款,合同附件明确包含\\\"全程跟拍记录\\\"服务项目。\\n(+5分) 婚礼当日,某摄影服务公司未经告知将摄像服务转包给文化传媒公司。\\n(+5分) 文化传媒公司工作室将录像全部丢失,未能交付原告。\\n(+ ... [TRUNCATED 762 chars] ... 人具有人身意义的特定物造成严重精神损害的,被侵权人有权请求精神损害赔偿。\\n(+5分)《最高人民法院关于确定民事侵权精神损害赔偿责任若干问题的解释》第五条\\n精神损害的赔偿数额根据以下因素确定:(一)侵权人的过错程度,但是法律另有规定的除外;(二)侵权行为的目的、方式、场合等具体情节;(三)侵权行为所造成的后果;(四)侵权人的获利情况;(五)侵权人承担责任的经济能力;(六)受理诉讼法院所在地的平均生活水平。\", \"points\": \"15\", \"tags\": \"法条依据得分\"}]",
|
||
"max_points": 60,
|
||
"prompt": "\n## 角色\n\n你是一名具有十年以上执业经验的法律实务专家,精通中国现行法律法规与司法实践。你擅长将复杂的法律问题分解为清晰的逻辑模块,并严格依据“结论先行、事实为重、推理严密、依据支撑”的专业风格进行解答。\n\n## 核心要求\n\n1. 严格顺序:回答必须按照以下四部分顺序展开,并使用对应标题:\n【结论】\n【案件事实】\n【推理过程】\n【法条依据】\n2. 内容规范:\n结论:直接、明确,针对提问的核心争议点给出肯定或否定的判断。\n案件事实:基于用户提供的案情,简明、客观地摘录与法律判断 ... [TRUNCATED 1912 chars] ... 并赔偿精神损害抚慰金。庭审中查明,某摄影服务公司已完成除摄像外的其他服务项目;某文化传媒公司系独立法人,其工作人员在操作设备时存在重大过失。另查,某甲在签订合同时未特别声明婚礼录像的重要性,但合同附件中列有\"全程跟拍记录\"服务项目。某摄影服务公司辩称其仅需承担合同违约责任,精神损害赔偿缺乏依据。某文化传媒公司以非合同相对方为由拒绝承担责任。\n\n## 问题\n以【结论 + 案情简述 + 分析过程+依据法条】的逻辑回答以下问题:在上述案例中,某甲能否向某摄影服务公司主张精神损害赔偿?\n"
|
||
}
|
||
}
|
||
```
|
||
|
||
*Note: Some content was truncated for display.*
|
||
|
||
## Prompt Template
|
||
|
||
**Prompt Template:**
|
||
```text
|
||
{question}
|
||
```
|
||
|
||
## Usage
|
||
|
||
### Using CLI
|
||
|
||
```bash
|
||
evalscope eval \
|
||
--model YOUR_MODEL \
|
||
--api-url OPENAI_API_COMPAT_URL \
|
||
--api-key EMPTY_TOKEN \
|
||
--datasets plawbench \
|
||
--limit 10 # Remove this line for formal evaluation
|
||
```
|
||
|
||
### Using Python
|
||
|
||
```python
|
||
from evalscope import run_task
|
||
from evalscope.config import TaskConfig
|
||
|
||
task_cfg = TaskConfig(
|
||
model='YOUR_MODEL',
|
||
api_url='OPENAI_API_COMPAT_URL',
|
||
api_key='EMPTY_TOKEN',
|
||
datasets=['plawbench'],
|
||
dataset_args={
|
||
'plawbench': {
|
||
# subset_list: ['case_analysis', 'legal_consultation', 'plaintiff_statement'] # optional, evaluate specific subsets
|
||
}
|
||
},
|
||
limit=10, # Remove this line for formal evaluation
|
||
)
|
||
|
||
run_task(task_cfg=task_cfg)
|
||
```
|