What it measures

Chinese-language academic and professional questions broaden knowledge evaluation beyond English.

How to read the result

Look at subject and language-specific errors rather than using the score as a translation of an English benchmark.

Limits and comparison traps

Language proficiency, culturally specific knowledge, and prompt conventions contribute jointly to the result.

Published evaluation

Macro average over all 67 subjects; source model class is retained.

Where this fits in an evaluation suite

Read accuracy together with error and abstention rates. Domain averages can hide language gaps, specialist failures, and poorly calibrated confidence.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Macro average over all 67 subjects; source model class is retained.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

  1. CMMLU: Measuring massive multitask language understanding in Chinese — published result table ↗

    Li, Haonan, Zhang, Yixuan, Koto, Fajri, Yang, Yifei, Zhao, Hai, Gong, Yeyun et al. · 2023

    All 8 authors

    Li, Haonan, Zhang, Yixuan, Koto, Fajri, Yang, Yifei, Zhao, Hai, Gong, Yeyun, Duan, Nan, Baldwin, Timothy

  2. Primary benchmark documentation or model evaluation report ↗

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.