WikiBench (OpenCompass)

OpenCompass Chinese four-way Wikipedia-style MCQ set, usually reported under circular evaluation rather than single-pass accuracy.

Also known as: opencompass/WikiBench, wikibench-wiki-single_choice_cn

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryknowledge
SubcategoryChinese four-way multiple-choice questions drawn from Wikipedia-style knowledge
Page statusunknown
Metriccircular accuracy (CircularEvaluator); AccEvaluator if circular is turned off
Directionhigher_is_better
Unit%
Dataset licenceApache-2.0 on OpenCompass code. Dataset licence is not stated on a public card (Hub page 404; datasets_info hf_id is empty).
PublisherOpenCompass (Shanghai AI Laboratory)

What it measures

OpenCompass WikiBench is a Chinese single-choice knowledge quiz. Each item is a question with four options A-D. Prompts and few-shot exemplars are encyclopaedic (biology, places, sports, Chinese history). The default configs enable circular evaluation: the same item is asked under four option rotations and the model must get every rotation right. This page is that OpenCompass dataset, not Wikipedia-article-quality "WikiBench" work from the HCI literature.

Task format

Four-way MCQ in Chinese. Three config styles exist: generative zero-shot (wikibench_gen.py imports wikibench_gen_f96ece.py), generative chain-of-thought (wikibench_gen_0978ad.py), and few-shot perplexity (wikibench_few_shot_ppl_c23d79.py). Generation post-processes with first_option_postprocess over ABCD. CircularEvaluator is on by default (do_circular = True). Dataset class WikiBenchDataset, path opencompass/WikiBench, file single_choice_cn.jsonl.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub