OpenCompass's Chinese-prompt 5-way CommonsenseQA wrap, scored as accuracy on a local validation.jsonl file.
unassessed
| Category | reasoning |
|---|---|
| Subcategory | Chinese-prompt 5-way commonsense multiple-choice (OpenCompass) |
| Page status | unknown |
| Metric | accuracy |
| Direction | higher_is_better |
| Unit | % |
| Publisher | OpenCompass (open-compass/opencompass) |
commonsenseqa_cn is OpenCompass dataset abbr commonsenseqa_cn, not English CommonsenseQA, not Bangla CommonsenseQA, and not CHARM. The model sees a Chinese 5-way question with options A-E and must pick the answer key. The loader CommonsenseQADataset_CN reads a local JSONL that already has a question field plus a HuggingFace-style choices.text list of five strings. Whether those questions are translations of English CommonsenseQA or a new Chinese set is not stated in the config, the loader, or OpenCompass's dataset-index (paper field empty).
Five-way multiple choice. Two OpenCompass configs: generation (GenInferencer, first_capital_postprocess, prompt ends with "答案：") and perplexity (PPLInferencer over the five choice texts). Both use ZeroRetriever (zero-shot) and AccEvaluator. Reader test_split is validation.
No model card in ModelSpec reports this benchmark yet.