kale667/interviewbench_cn

---
language:
- zh
license: other
pretty_name: InterviewBench-CN
size_categories:
- 100K<n<1M
task_categories:
- question-answering
- text-generation
- text-classification
tags:
- chinese
- interview
- benchmark
- multi-task
- evaluation
configs:
- config_name: clean
default: true
data_files:
- split: train
path: clean_264178.jsonl
- config_name: answer_generation
data_files:
- split: train
path: task_answer_generation.jsonl
- config_name: answer_evaluation
data_files:
- split: train
path: task_answer_evaluation.jsonl
- config_name: followup_generation
data_files:
- split: train
path: task_followup_generation.jsonl
- config_name: error_diagnosis
data_files:
- split: train
path: task_error_diagnosis.jsonl
- config_name: multiturn_interview
data_files:
- split: train
path: task_multiturn_interview.jsonl
- config_name: gold_review
data_files:
- split: train
path: gold_review_queue_10000.jsonl
---

# InterviewBench-CN

InterviewBench-CN 是一个面向中文求职与技术面试场景的大规模、多任务、可追踪交互式面试基准。

原始数据共 381,945 条，来自 Hugging Face 中文面试数据集及 300 余个非 HF 公开来源。经过格式统一、噪声过滤、精确去重、归一化去重和语义近重复清洗后，形成 264,178 条可追踪 clean 样本。

本数据集支持五类任务：

- Answer Generation：根据面试问题生成标准回答
- Answer Evaluation：根据问题和候选回答输出 1-5 分评分及反馈
- Follow-up Generation：根据回答生成进一步追问
- Error Diagnosis：识别错误回答中的遗漏、误解、幻觉和浅层回答
- Multi-turn Interview：基于多轮历史生成下一问和总评

所有样本均保留来源站点、URL、文件路径和源内位置，支持溯源和复核。