choucsan/commonsense_qa_zh

Commonsense QA Chinese Multiple-Choice Dataset

This dataset is a Chinese four-choice SFT version of tau/commonsense_qa. It is designed to supplement commonsense multiple-choice training data for benchmark tasks such as challenge_common_sense.

The original dataset is in English and contains five-choice commonsense questions. This release keeps only samples that can be aligned to the official four-choice benchmark format, translates the question and options into Chinese, and formats each answer as a direct /no_think response.

---

Pipeline

The dataset is built through a filtering, translation, and verification pipeline:

1. Source loading: We read question, choices, and answer_key from the original tau/commonsense_qa dataset.
2. Four-choice alignment: We keep only samples whose correct answer is in A-D and whose A/B/C/D options are complete. Original samples requiring option E are excluded.
3. Chinese translation: We use Qwen3-14B to translate the English question and answer options into Chinese.
4. Label preservation: The response field is constructed directly from the original answer_key. The answer label is not generated or changed by Qwen3-14B.
5. Quality filtering: We filter samples with invalid format, missing options, JSON parsing failures, or translated prompts that expose the correct answer.
6. Benchmark formatting: We convert each retained sample into the same prompt and response style used by the official challenge_common_sense benchmark, including the /no_think suffix.

---

Dataset Structure

commonsense_qa_zh/
├── commonsense_qa_zh.jsonl      # Chinese four-choice SFT samples
└── README.md                    # Dataset card

Each line in commonsense_qa_zh.jsonl is one training sample. The sample is stored as a list containing one dialogue-style object:

[
  {
    "system": "...",
    "prompt": "...",
    "response": "..."
  }
]

---

Dataset Overview

• Source dataset: tau/commonsense_qa
• Processed file: commonsense_qa_zh.jsonl
• Total samples: 8,761
• Language: Chinese prompts translated from English source questions
• Task type: Commonsense multiple-choice question answering
• Choice format: Four options, A-D
• Response format: 正确答案是 X
• Thinking mode: /no_think
• Translation model: Qwen3-14B
• Answer source: Original answer_key, not model-generated text

commonsense_qa_zh.jsonl

Field	Type	Description
system	string	System prompt used for the sample
prompt	string	Chinese question, four options A-D, benchmark answer instruction, and /no_think suffix
response	string	Ground-truth answer in the fixed format 正确答案是 X

Example:

[
  {
    "system": "你是一个非常聪明的助手，请直接遵循指示作答。",
    "prompt": "请回答以下问题：\n\n对学校的制裁是一次沉重的打击，它们似乎要什么学校为改变所做的努力？\nA.忽视\nB.执行\nC.专制的\nD.大声吼叫\n\n请按以下格式作答：\"正确答案是 (在此处填写选项字母)\"/no_think",
    "response": "正确答案是 A"
  }
]

---

Intended Use

This dataset can be used directly as SFT data for Chinese commonsense multiple-choice QA. It is especially useful when the target evaluation requires the model to answer in the official benchmark format:

正确答案是 X

The /no_think suffix is intentionally included in every prompt, so the expected behavior is to output the final option answer directly rather than generate an intermediate reasoning trace.

---

Notes

• This is a translated and benchmark-aligned derivative dataset, not a full mirror of the original tau/commonsense_qa dataset.
• Original five-choice samples whose correct answer is E are not included.
• Qwen3-14B is used for translation and filtering assistance, while answer labels remain tied to the original dataset.
• The dataset is intended for supervised fine-tuning and benchmark-format alignment, not for evaluating raw English CommonsenseQA performance.

---

Contact

For questions, corrections, or collaboration requests:

choucisan@gmail.com