Reasoning Benchmark

C³

C3 - Multiple-Choice Chinese Machine Reading Comprehension dataset (CLUE benchmark family; test items linked from github.com/CLUEbenchmark/CLUE). 0-shot accuracy as reported on the Tencent Hy4-flash card.

Category

Reasoning

Score type

percent

Scale

100%

higher is better

Models tested

2

What is C³?

C³ is a benchmark for general reasoning and comprehension. The official description says: "C3 - Multiple-Choice Chinese Machine Reading Comprehension dataset (CLUE benchmark family; test items linked from github.com/CLUEbenchmark/CLUE)". Below we describe in simple words what benchmarks of this type generally examine.

What does it test?

C³ measures general reasoning and comprehension. In detail, it checks:

  • drawing the right conclusion from a passage or scenario
  • commonsense judgement about everyday situations and words
  • counting, comparing and reasoning over several facts at once
  • avoiding shortcuts that memorise answers instead of reasoning

How to read the score

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Official methodology note: C3 - Multiple-Choice Chinese Machine Reading Comprehension dataset (CLUE benchmark family; test items linked from github.com/CLUEbenchmark/CLUE). 0-shot accuracy as reported on the Tencent Hy4-flash card.

Frequently asked questions

What is the C³ benchmark?

C³ is a benchmark for general reasoning and comprehension. The official description says: 'C3 - Multiple-Choice Chinese Machine Reading Comprehension dataset (CLUE benchmark family; test items linked from github.com/CLUEbenchmark/CLUE)'. Below we describe in simple words what benchmarks of this type generally examine.

What does C³ test?

C³ mainly measures general reasoning and comprehension. Concretely, it checks: drawing the right conclusion from a passage or scenario; commonsense judgement about everyday situations and words; counting, comparing and reasoning over several facts at once.

How is the C³ score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get C³ data and scores?

The official source is https://github.com/CLUEbenchmark/CLUE. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/c3/scores) or pull it directly from the Open LLM Radar API with any HTTP client.

Model scores 2 models tested

Reset
# Model Score Date Source Verified
1 DeepSeek-V4-Pro
DeepSeek
78.3%
— Tencent Hy4-flash card few-shot pretrain eval, 'DeepSeek-V4-Pro-Base' column, 0-shot (https://huggingface.co/tencent/Hy4-flash); Tencent's own eval of this base-variant artifact - not self-reported Unverified
2 DeepSeek-V4-Flash
DeepSeek
74.1%
— Tencent Hy4-flash card few-shot pretrain eval, 'DeepSeek-V4-flash-Base' column, 0-shot (https://huggingface.co/tencent/Hy4-flash); Tencent's own eval of this base-variant artifact - not self-reported Unverified

What now?

You've seen how models score on C³ — here's what to do next.