Chinese two-choice value-alignment eval whose public OpenCompass slice is 1,712 responsibility items that ask which of two replies is more responsible.
unassessed
| Category | safety |
|---|---|
| Subcategory | Chinese human-values alignment (safety vs responsibility multiple-choice) |
| Page status | active |
| Metric | accuracy |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 1712 |
| Dataset licence | Apache-2.0 |
| Publisher | Alibaba Group (X-PLUG) and Beijing Jiaotong University |
CValues tests whether a Chinese LLM prefers the safer or more responsible of two replies to a sensitive prompt. The paper splits the skill into safety (level-1: refuse harm) and responsibility (level-2: refuse and still give a constructive answer). OpenCompass, which is how this id is wired in this repository, ships only the public responsibility multiple-choice slice. The model sees one Chinese question and two replies labelled A and B after a rewrite of 回复1/回复2, and must name the better reply.
Two-way multiple choice in Chinese. OpenCompass generates a free-text answer, then first_capital_postprocess plus AccEvaluator score the A/B letter. Zero-shot (ZeroRetriever). The authors' own script instead parses 回复1/回复2 and reports accuracy both with and without refusals.
No model card in ModelSpec reports this benchmark yet.